UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Integrative approaches to environmental life cycle assessment of consumer electronics and connected media Teehan, Paul 2014

You don't seem to have a PDF reader installed, try download the pdf

Item Metadata


24-ubc_2014_september_teehan_paul.pdf [ 2.27MB ]
JSON: 24-1.0167496.json
JSON-LD: 24-1.0167496-ld.json
RDF/XML (Pretty): 24-1.0167496-rdf.xml
RDF/JSON: 24-1.0167496-rdf.json
Turtle: 24-1.0167496-turtle.txt
N-Triples: 24-1.0167496-rdf-ntriples.txt
Original Record: 24-1.0167496-source.json
Full Text

Full Text

Integrative approaches to environmental life cycle assessment ofconsumer electronics and connected mediabyPaul TeehanM.A.Sc., University of British Columbia, 2008B.A.Sc., University of Waterloo, 2006A THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFDoctor of PhilosophyinTHE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES(Resource Management and Environmental Studies)The University of British Columbia(Vancouver)June 2014c© Paul Teehan, 2014AbstractThe environmental impacts of information and communication technologies and consumer electronicsare challenging to evaluate. Organizations and individuals wishing to reduce the impacts attributableto their usage of these products and systems rely on a limited technical knowledge base that strugglesto stay current. Using a life cycle assessment approach which expresses environmental impacts quan-titatively in terms of greenhouse gas emissions and primary energy demand, this dissertation signifi-cantly expands our understanding of the impacts of desktop computers, electronics products in general,and connected media services accessed in the home, in order to support environmentally-consciousdecision-making and policy regarding these products and systems.The first of three studies, a meta-analysis of prior life cycle assessments of desktop PCs, resolves animportant ambiguity in this literature and demonstrates that greenhouse gas emissions due to operationalenergy consumption usually exceed those due to device manufacturing. The second study calculatesembodied greenhouse gas emissions of eleven electronics products through a teardown analysis, andfinds a linear relationship between mass and embodied emissions, thus demonstrating that lightweight,compact products offer environmental benefits relative to larger products. A comparison to studies ofolder products also reveals that newer products are more materially efficient, largely due to reductionin integrated circuit content per product. Finally, the third study calculates aggregate US consumergreenhouse gas emissions due to broadcast television, video on demand, online video, other onlineuses, and offline uses when consumed using televisions, personal computers, tablets, and smartphones,including emissions due to devices in the home, networks, and datacenters. The study concludes thatemissions due to video end-uses account for 75% of total consumer ICT emissions. About 71% ofconsumer ICT emissions arise due to devices in the home, especially TVs and desktop PCs, with theremainder due to networks and datacenters. Mobile platforms using Wi-Fi connections are the leastiiimpactful mode of consuming connected media content.Collectively, the dissertation argues for a more integrated approach towards impact estimation, inorder to surmount issues regarding variation of modeling assumptions across existing studies, longevityof published work, and coverage of emerging products and services.iiiPrefaceThe research chapters in this dissertation were written as manuscripts and are each either published inpeer-reviewed journals, or intended to be submitted for publication.Chapter 3 was originally published as “Sources of variation in life cycle assessments of desktopcomputers”, authors Paul Teehan and Milind Kandlikar [1], in the Journal of Industrial Ecology, 2012.I conceived of the research, conducted the literature review, performed the analysis, and wrote about95% of the manuscript. Dr. Kandlikar helped in framing and structuring the manuscript. The copyrightto this article is held by the journal publisher, Yale University; republication within this dissertation ispermitted under the license terms. The original article has been edited to conform to the dissertation’sstyle regarding table and figure numbering, citations and references, and units. Appendix A contains thesupplementary material originally published with the article, edited slightly to conform to dissertationstyle.Chapter 4 was originally published as “Comparing embodied greenhouse gas emissions of moderncomputing and electronics products”, authors Paul Teehan and Milind Kandlikar [2], in EnvironmentalScience & Technology, 2013. I conceived of the research and designed the study, with some assistancefrom Drs. Kandlikar, Tony Bi, and Hadi Dowlatabadi, who provided critical feedback in the study’s con-ceptual stages. I conducted 100% of the study, including the teardown analysis, life cycle assessment,and statistical analysis, and wrote about 95% of the manuscript; Dr. Kandlikar provided comments andedits to the manuscript. The statistical analysis was not present in the initial manuscript submissionand was suggested by the journal’s anonymous reviewers. The copyright to this article is held by thejournal publisher, the American Chemical Society (ACS); republication within this dissertation is per-mitted under the license terms. The original article has been edited to conform to the dissertation’s styleregarding table and figure numbering, citations and references, and units. The original article’s sup-ivplementary material is available online under an open access license through the publisher’s website1.Appendix B contains an edited version of the supplementary material which has been re-organized toimprove flow and expanded slightly to improve accessibility for non-specialist readers; extensive labo-ratory measurements provided in the original supplementary material are not included in Appendix Bas they are lengthy and not integral to the communication of the research methods and results.A preliminary and much smaller version of the study in Chapter 5 was published as “Estimatingthe changing environmental impacts of ICT-based tasks: a top-down approach”, authors Paul Teehan,Milind Kandlikar, and Hadi Dowlatabadi [3], in the conference proceedings of the International Sympo-sium on Sustainable Systems and Technology, 2010, which are not peer-reviewed. The research aims ofthe preliminary study were initially conceived of by myself and developed through collaborative discus-sions with Drs. Kandlikar and Dowlatabadi. The present manuscript in Chapter 5 has similar high-levelgoals to the preliminary study but was re-done from scratch with expanded scope and improved methodsand data sources. This manuscript is intended to be submitted for publication in a peer-reviewed jour-nal with authors Paul Teehan, Eric Masanet, and Milind Kandlikar. I designed the methods, compiledthe input data, performed 100% of the analysis, and wrote about 95% of the manuscript. Dr. Masanetprovided technical feedback on a draft of the study and suggested some improvements to the methodsand input data sources. Drs. Masanet and Kandlikar each provided editorial feedback relating to theframing and structure of the manuscript.1http://pubs.acs.org/doi/abs/10.1021/es303012rvTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiiDedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiv1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Research approach and goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31.2 Significance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51.3 Thesis overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 Context and background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82.1 Conceptualizing impacts of ICT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82.1.1 The three-level taxonomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82.1.2 Economic frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102.2 LCA and ICT products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142.2.1 Methodological approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . 142.2.2 Industry profile: On Moore’s Law and energy . . . . . . . . . . . . . . . . . . 192.2.3 Knowledge gaps and research approach . . . . . . . . . . . . . . . . . . . . . 202.3 LCA and ICT services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.3.1 Methodological approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.3.2 Industry profile: connected media and the cloud . . . . . . . . . . . . . . . . . 252.3.3 Knowledge gaps and research approach . . . . . . . . . . . . . . . . . . . . . 262.4 Methodological limitations and justification . . . . . . . . . . . . . . . . . . . . . . . 292.4.1 LCA database and LCIA impact scheme uncertainties . . . . . . . . . . . . . 302.4.2 Truncation error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332.4.3 Toxicity and environmental health . . . . . . . . . . . . . . . . . . . . . . . . 352.4.4 General bounds on LCA results . . . . . . . . . . . . . . . . . . . . . . . . . 362.4.5 Alternative approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372.5 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38vi3 Meta-analysis of LCAs of desktop PCs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393.2 Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403.2.1 Scope and unit of analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403.2.2 Definition of environmental impact . . . . . . . . . . . . . . . . . . . . . . . 413.2.3 Analytical approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413.3 List of studies and their overall results . . . . . . . . . . . . . . . . . . . . . . . . . . 423.4 Manufacturing and production phases . . . . . . . . . . . . . . . . . . . . . . . . . . 463.4.1 Component-level impacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473.4.2 Integrated circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483.4.3 Total manufacturing impact: analysis . . . . . . . . . . . . . . . . . . . . . . 493.5 Use phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503.5.1 Unit energy consumption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503.5.2 Product lifespan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523.5.3 Total use phase impact: analysis . . . . . . . . . . . . . . . . . . . . . . . . . 533.6 Analysis of overall impact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 574 Embodied emissions of ICT devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 594.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 594.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 614.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 634.3.1 Product composition by mass . . . . . . . . . . . . . . . . . . . . . . . . . . 634.3.2 Embodied emissions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 634.3.3 Comparison with other studies . . . . . . . . . . . . . . . . . . . . . . . . . . 654.3.4 Data quality and uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . 674.4 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 674.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 715 Emissions due to connected media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 735.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 735.1.1 Study scope and boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 755.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 765.2.1 Device emissions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 775.2.2 Network emissions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 785.2.3 Datacenter emissions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 805.2.4 Overall emissions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 805.3 Model inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 815.3.1 Device and platform characteristics . . . . . . . . . . . . . . . . . . . . . . . 815.3.2 Network and datacenter characteristics . . . . . . . . . . . . . . . . . . . . . 825.3.3 End-use data traffic and time spent . . . . . . . . . . . . . . . . . . . . . . . . 845.4 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 865.4.1 Emissions by platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 865.4.2 Emissions by end-use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 875.4.3 Uncertainty analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 885.4.4 Energy comparison to household appliances . . . . . . . . . . . . . . . . . . . 945.5 Implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95vii6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 986.1 Specific findings and significance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 986.1.1 Life cycle assessments of desktop PCs . . . . . . . . . . . . . . . . . . . . . . 986.1.2 Embodied emissions of electronics . . . . . . . . . . . . . . . . . . . . . . . . 996.1.3 Emissions of connected media . . . . . . . . . . . . . . . . . . . . . . . . . . 1006.2 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1016.2.1 General limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1016.2.2 LCA meta-analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1026.2.3 Product LCA and embodied emissions . . . . . . . . . . . . . . . . . . . . . . 1026.2.4 Connected media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1036.3 Future research needs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1036.3.1 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1036.3.2 ICT products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1046.3.3 ICT services and infrastructure . . . . . . . . . . . . . . . . . . . . . . . . . . 1056.4 Implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1066.4.1 Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1066.4.2 Commentary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1076.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110A Supplementary material for Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130A.1 Summary data tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130A.2 Mapping to subassembly categories . . . . . . . . . . . . . . . . . . . . . . . . . . . 134A.3 Integrated circuit content . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135B Supplementary material for Chapter 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137B.1 Summary data tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137B.2 Adjustments and comparison to ecoinvent . . . . . . . . . . . . . . . . . . . . . . . . 140B.2.1 System boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141B.2.2 Silicon die and integrated circuits . . . . . . . . . . . . . . . . . . . . . . . . 141B.2.3 LCD assembly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141B.2.4 LCD power supply . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142B.2.5 Other variation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142B.3 Uncertainty factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143B.4 Silicon die content in integrated circuits . . . . . . . . . . . . . . . . . . . . . . . . . 144B.4.1 Discrepancy in ecoinvent models . . . . . . . . . . . . . . . . . . . . . . . . . 146B.4.2 Stacked IC’s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147B.5 Linear regression model selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148C Supplementary material for Chapter 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151C.1 Summary data tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151C.2 Model input parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155C.2.1 Device parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155C.2.2 Network and datacenter intensities . . . . . . . . . . . . . . . . . . . . . . . . 157C.2.3 End-use parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163C.3 Uncertainty model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168C.3.1 Parameter uncertainty estimation . . . . . . . . . . . . . . . . . . . . . . . . . 168viiiList of TablesTable 2.1 Review of studies of ICT end-uses . . . . . . . . . . . . . . . . . . . . . . . . . . 27Table 2.2 Selected LCIA results for selected processes modeled in ecoinvent V2 database . . 31Table 2.3 Ratio of modeled outputs from GaBi 4 relative to ecoinvent v2, selected processesand impact categories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32Table 3.1 IC inventories and impacts per desktop mainboard, with originally reported inventoryand adjusted inventory assuming 0.2g/cm2 finished die . . . . . . . . . . . . . . . . 48Table 4.1 Products analyzed in this study . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60Table 4.2 Coefficients and statistics from model fitting . . . . . . . . . . . . . . . . . . . . . 70Table 5.1 Device and platform assumptions for operational energy, embodied GHG emissions,and US consumer installed base . . . . . . . . . . . . . . . . . . . . . . . . . . . 81Table 5.2 Estimated US monthly consumer data traffic per platform instance across fixed andmobile networks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85Table 5.3 Estimated monthly data traffic DEU,P and time spent TEU,P per platform instance, foreach end-use and platform. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85Table 5.4 Estimated confidence intervals on model input parameters. . . . . . . . . . . . . . . 93Table A.1 Data table for Figure 3.1: Total impacts by life cycle phase . . . . . . . . . . . . . 130Table A.2 Data table for Figure 3.2: Manufacturing impacts by subassembly . . . . . . . . . . 131Table A.3 Data table for Figure 3.3: Mass and impact factors by subassembly . . . . . . . . . 131Table A.4 Data table for Figure 3.4: Desktop unit energy . . . . . . . . . . . . . . . . . . . . 132Table A.5 Data table for Figure 3.5: Desktop lifespan . . . . . . . . . . . . . . . . . . . . . . 133Table A.6 Data table for Figure 3.6: Electricity impact factors . . . . . . . . . . . . . . . . . 133Table B.1 Summary data table, product composition by mass . . . . . . . . . . . . . . . . . . 137Table B.2 Summary data table, embodied GHG emissions, in kg CO2e . . . . . . . . . . . . . 138Table B.3 Summary data table, cumulative energy demand, in MJ . . . . . . . . . . . . . . . 138Table B.4 Summary data table, mass, embodied GHG, and embodied primary energy demand,this study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139Table B.5 Summary data table, mass and embodied GHG, Apple dataset . . . . . . . . . . . . 139Table B.6 Global warming potential results for original ecoinvent study and this study’s adjust-ments, in kg CO2e . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140Table B.7 Pedigree matrix scoring for uncertainty characterization: assessment of ecoinventcomponent models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144Table B.8 Monte Carlo analysis results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145Table B.9 Model selection results for the top fifteen models. . . . . . . . . . . . . . . . . . . 149ixTable B.10 Linear regression model outputs, this study’s dataset . . . . . . . . . . . . . . . . . 150Table B.11 Linear regression model outputs, Apple’s dataset . . . . . . . . . . . . . . . . . . . 150Table C.1 Total US consumer emissions by end-use and location, with Monte Carlo results . . 151Table C.2 Energy and GHG emissions by end-use and platform . . . . . . . . . . . . . . . . . 153Table C.3 Energy and GHG emissions by platform . . . . . . . . . . . . . . . . . . . . . . . 154Table C.4 Device operational energy assumptions, with 2006 baseline from [4] and 2010 base-line from [5] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155Table C.5 Total traffic from all devices, comparing this study assumptions with original Ciscomodel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164Table C.6 Derivation of time spent with online video . . . . . . . . . . . . . . . . . . . . . . 165Table C.7 Modal share of time spent online per person . . . . . . . . . . . . . . . . . . . . . 167xList of FiguresFigure 2.1 Illustration of impact trajectories under consumption Y and technology T . . . . . 11Figure 3.1 Summary of LCA studies showing breakdown of life cycle energy use and carbonequivalent impacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45Figure 3.2 Manufacturing impacts of desktop control unit components (relative) . . . . . . . 46Figure 3.3 Mass and impacts of desktop PC components (excluding mainboard ICs) . . . . . 47Figure 3.4 Review of unit energy consumption of desktop PCs . . . . . . . . . . . . . . . . 51Figure 3.5 Review of lifespan of desktop PCs . . . . . . . . . . . . . . . . . . . . . . . . . 53Figure 3.6 Impact factors for electricity from LCA studies and ecoinvent database . . . . . . 54Figure 3.7 Overall primary energy and global warming potential, showing prior results and thisstudy’s estimates of reasonable ranges for a typical desktop PC without display . . 56Figure 4.1 Results showing product mass (a) and embodied emissions (b). (ei) denotes adjustedstudies from ecoinvent database. . . . . . . . . . . . . . . . . . . . . . . . . . . . 64Figure 4.2 Monte Carlo results: mean embodied GHG emissions with error bars showing ±two standard deviations, using data quality pedigree matrix approach . . . . . . . . 68Figure 4.3 Residuals from model selection . . . . . . . . . . . . . . . . . . . . . . . . . . . 69Figure 5.1 Block diagram showing study scope and boundary . . . . . . . . . . . . . . . . . 75Figure 5.2 Network and datacenter energy, emissions, and intensities (note log scales). . . . . 83Figure 5.3 Emissions per platform and US consumer total, for all platforms, including devices,networks, and datacenters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87Figure 5.4 Emissions per platform, by end-use . . . . . . . . . . . . . . . . . . . . . . . . . 88Figure 5.5 Emissions across all platforms, by end-use and overall, US consumer total . . . . 89Figure 5.6 Emissions across all end-uses, by platform and overall, US consumer total . . . . 89Figure 5.7 Emissions per platform and US consumer total, baseline and efficient datacenterscenarios. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92Figure 5.8 Monte Carlo results showing uncertainty range for GHG emissions, US consumertotal; error bars span ± two standard deviations. . . . . . . . . . . . . . . . . . . . 94Figure 5.9 Comparison of ICT energy use against other electric household appliances . . . . 95Figure B.1 Silicon die area measurements via X-ray, with regression lines . . . . . . . . . . . 146xiAcknowledgmentsThis research was primarily supported by the National Sciences and Engineering Research Councilof Canada (NSERC), the Pacific Institute for Climate Solutions (PICS), and UBC’s Bridge Program.Scientific software used in conducting research in this dissertation was funded by the UBC SustainabilityInitiative. The study in Chapter 4 was additionally supported through in-kind donations from DellCanada and from The Hackery in Vancouver; open access publishing for this study was sponsored by theCenter for Climate and Energy Decision-Making at Carnegie Mellon University (NSF SES-0949710).Sincere thanks to my supervisor, Prof. Milind Kandlikar, for accompanying me on the journey ofthe last six years with patience and kindness, and to Prof. Hadi Dowlatabadi, for challenging me tobetter myself. I would like to acknowledge the invaluable technical assistance from Prof. Eric Masanet,my collaborator in the study in Chapter 5, which greatly elevated this work. In addition, I benefitedfrom the expertise of Profs. Kandlikar and Dowlatabadi, as well as Prof. Tony Bi, and my graduatestudent colleagues at UBC, especially Stefan Storey, Matt Dolf, and Rob Sianchuk. Finally, sincerethanks to the administrative staff at the Institute for Resources, Environment and Sustainability for theircontinuing hard work and support.I’ve been priviledged to share the space at AERL with so many bright, engaged, genuine, and kindpersonalities, many of whom I am fortunate to count as friends: Stefan Storey, Bertine Stelzer, CynthiaMorinville, Lisa Westerhoff, Tee Lim, Maryam Rezai, Matt Dolf, Joshua Welsh, James Murphy, ElaineHsiao, Christian Beaudrie, Olivia Freeman, Laura Cornish, Jordan Tam, Gerald Singh, Julia Freeman,Conor Reynolds, Francies Ries, Brian Gouge, Sara Elder, Ed Gregr, Sarah Klain, and all the others.Special thank you to the afternoon writing club, especially Stefan, Matt, and Maryam, for being withme in the final push.Many thanks to my friends who have been with me during the last six years, festival camp-mates,xiiband-mates, brunch-mates, everyone who was close enough to hug – Eric Jacobsen, Lisa Rockwell,Mike Gevaert, Kristjan Bergey, Kristen Walmsley, Marielle Camozzi, Kerry Freek, Gerrit van Wouden-berg, Matt Siegle, Aleeza Gerstein, Kay and Kristin, Xianny Ng, Kerry Nickols, Kurt Vaughan, ClaireRay, James McConaghie, Hande Gungor, Stan Iordanov, Aaron Colin, Steve Luscher, Kate Rossiter,Kyle Sulyma, Bryan Chiu. At this stage, I believe one can count success in life by the frequency andduration of warm, loving hugs. Thanks to my brothers, whom I look forward to seeing more of on theother side, and to my parents, who have done a superb job in deliberately not asking those forbiddenquestions about progress.In many ways the completion of this program represents the beginning of a new life chapter whichpromises to be rich, colourful, and alive, in part because of the vision and perspective I’ve gained overthese years. There are three people in particular who were instrumental in helping me achieve thisvision, without whom my task would have been immeasurably harder. First, Kevin Tate, my supervisorat Pulse Energy, has been incredibly supportive and flexible throughout the last year, far beyond myexpectations; I greatly look forward to the coming years at Pulse. Second, Dr. Douglas Ozier, mycounselor, has helped me become the person I’ve always wanted to be; our work has been transformativeand life-changing in ways that are hard to overstate. Finally, Rune McKellar, my fiance´e, has shown methat sometimes dreams can come true; that it is okay to believe in love, peace, and truth; and that therecan be magic in the world.xiiiFor Rune,who gave me myself.xivChapter 1IntroductionThe research in this dissertation evaluates the environmental impacts related to information and commu-nication technologies and consumer electronics (ICTs), including computers, servers, mobile electronicdevices, TVs, and related devices. ICTs impact the environment through many mechanisms, includingthrough operational energy consumption, through material and energy inputs and pollutant outputs inmanufacturing and disposal processes, and through changes in human behavior which may affect otherindustries. This work considers environmental impacts in terms of primary energy consumption andgreenhouse gas emissions (GHGs). The high-tech sector is fast-moving, characterized by rapid emer-gence and uptake of new products and services and corresponding change in the nature of environmentalimpacts due to ICT.These impacts have been described through several narratives, which have evolved over time. Growthof e-commerce and the internet in the 1990s was initially cast as a means of achieving “weightless”economic growth [6, 7, 8] leading to a more dematerialized society with reduced resource consump-tion [9, 10, 11]. A contrary response emerged with studies of the environmental impacts of electronics,notably the pioneering life cycle assessment-based work of Eric Williams [12, 13] and others, investiga-tive reports of social and environmental justice and public health issues surrounding informal e-wasterecycling [14, 15, 16], and alarmist forecasts predicting explosive growth in energy consumption due tothe internet [17] which were later debunked [18]. Amidst continuing research in each of these strands- see reviews on life cycle assessments [19, 20], impacts due to e-waste [21, 22], and studies of energyconsumption of the ICT sector [5, 23, 24] - a more pragmatic re-framing of earlier “weightless” narra-tives acknowledges the environmental downsides of the ICT industry, estimated to account for about 2 to13% of global GHG emissions [25] and about 13% of US residential and commercial building energy con-sumption [26], but argues that the application of ICT could improve efficiency and resource consumptionacross the economy, targeting the remaining 97% of GHG emissions [27, 28, 29, 30, 31, 32, 33], thoughthese studies are largely conjectural rather than predictive.Reducing overall impacts, whether considering impacts due to ICTs themselves or due to their appli-cation, requires identifying key leverage points at which interventions may be targeted, and supportingthe design and implementation of such interventions. There are many dimensions through which im-pacts due to ICTs may be considered, and thus many different types of interventions which have beenhistorically applied, such as efficiency standards, voluntary labeling schemes, guidance and informa-tion programs, restrictions on hazardous substances, and others. The goals of an intervention dependon those of the actor carrying it out; for example, firms may enact interventions in order to reduce theirown costs, while governments and non-profit advocacy groups may act to reduce environmental impactsthat arise as market externalities and thus will not normally be acted upon through market incentives.Likewise, the mechanism of an intervention depends on the capabilities of the actor; firms and individ-uals usually act only upon themselves, but government agencies can affect the market directly throughregulation, and indirectly through incentives and information programs.Scientifically rigorous quantifications of impacts play an important role in green decision-makingand policy-making, especially for interventions that are designed to reduce impacts that are normallyspecified quantitatively. For ICT-using firms and individuals, the ability to preferentially support lower-impact products, systems, and behaviors pre-supposes the existence of an operationally useful definitionof “lower-impact”. In the case of products, labeling schemes like ENERGY STAR [34] and EPEAT [35]which incorporate fundamental scientific research seek to provide this definition, though they have lim-itations, such as limited coverage of newer product families. For systems and behaviors that are consid-erably more complex, guidance is more limited.The research in this dissertation is designed to support decision-making among ICT-using firms andindividuals, both directly through the study of ICT behaviors, and indirectly through contributions to thefundamental scientific literature describing impacts of ICT products. The research findings also suggestpriorities for policy approaches to encourage green decision-making relating to ICT. Impacts are definedquantitatively in terms of energy use and GHG emissions using a life cycle assessment (LCA) approach.21.1 Research approach and goalsLife cycle assessment is an engineering methodology which models products and systems as linearsums of component processes, expressing the impacts of each component in quantitative terms, andaggregating component impacts to determine the overall impact of the product or system. LCA hasbecome a primary methodology for the quantitative determination of environmental impacts due to ICTproducts and services and has been applied to calculate impacts of products like desktop PCs [13],laptops [36], and monitors [37], and services like reading e-books [38], downloading music [39], andstreaming video [40], among others. Chapter 2 introduces LCA and reviews its prior application to ICTproducst and services.LCA studies are, by their nature, highly sensitive to modeling assumptions, such as geographic andtemporal scope, method of aggregation, source of industrial process data, and especially the definitionof the product or system under study. For this reason, LCAs are most effectively used either as a tool forcomparing the impacts of different stages of an industrial process, or for comparing the impacts of alter-native means of achieving a function; within the context of a study, modeling assumptions can be heldconstant allowing for valid comparisons. Comparisons of the results of one LCA study to another areonly valid if the studies are carefully adjusted to ensure consistency of modeling assumptions and inputdata, but carrying out such adjustments is time-consuming and sometimes impossible as the assump-tions may be incompletely specified and input data sources proprietary. Nevertheless, it is unrealistic toexpect researchers and analysts to start from scratch every time, especially when modeling the impactsof network-driven services which touch many different ICT products and systems including the internetitself; accordingly, the integration of prior results from multiple LCA studies is a standard practice asdisplayed in several recent studies, e.g. [39, 40, 41, 42, 43, 44, 45]. The knowledge base upon whichstudies of the impacts of complex ICT systems are built is always imperfect and often out of date, yeta research need to understand these impacts in accurate terms remains. Thus, a fundamental challengeis one of integration, which is hindered by opaqueness in modeling assumptions and source data fromexisting studies, and by the fast rate of change in the high tech sector that can outpace the academicliterature.ICT is a dynamic, complex field, characterized by rapid emergence and uptake of new products andservices which can quickly become embedded throughout societies and economies. Global demand for3ICT products and services is undoubtedly rising, which increases the need to understand the relationshipbetween these products and services and environmental impacts. Fundamental research exploring thisrelationship is hampered by many challenges, three of which are addressed in this dissertation. First, theinterpretation of LCA results is difficult due to the complexity of products being studied and variabilityin the ways in which they may be used (Chapter 3). Second, ongoing rapid proliferation and obsoles-cence of devices creates significant knowledge gaps (Chapter 4). Third, the rise of connected “cloud”services greatly increases the complexity of typical ICT end-uses (Chapter 5). Actionable policies tar-geting green decision-making rely on the precise information which is best obtained through detailedLCA, but responding to the methodological challenges presented by the dynamic ICT sector requires asimultaneous higher-level view, in order to anchor detailed LCA studies within the broader context.Given this need, all three studies in this dissertation are integrative in nature. The first study, pre-sented in Chapter 3, is driven by the observation that several prominent LCA studies of desktop PCshave contradictory results regarding the magnitude of GHG emissions attributable to the device, andin particular regarding whether the production or operational phases of the product life cycle are thelargest contributor to GHG emissions. Using meta-analysis of existing studies, this study addresses thefollowing research questions:1. Among existing published LCA studies of desktop PCs, what factors contribute to their disagree-ment regarding overall impacts and the relative contributing share of product life cycle phases?2. What can be concluded from existing LCA studies regarding the magnitude of GHG emissionsand cumulative energy demand, and the relative contribution of each phase of the product lifecycle to these indicators?The second study, presented in Chapter 4, is motivated by lack of coverage of new products inexisting LCA literature, difficulty in making product-by-product comparisons using existing literaturedue to variations in modeling assumptions across studies, and the challenge of adapting the existingknowledge base to assess new products. Using an LCA approach supported by teardown analysis of 11products, this study addresses the following research questions regarding embodied GHG emissions ofconsumer electronics:1. How do embodied emissions of modern devices compare to equivalent comparable devices fromolder product generations?42. How do embodied emissions of small mobile devices (e.g. tablet) and small form-factor devices(e.g. lightweight desktop) compare to those of larger devices (e.g. full-sized desktop)?3. To what extent may first-order estimates of embodied emissions be generated using limited prod-uct information, instead of conducting full LCA studies?The final study, presented in Chapter 5, is motivated by significant knowledge gaps and methodolog-ical challenges relating to the GHG emissions of emerging connected media ICT end-uses which arequickly replacing older modes of ICT usage. Using an integrated model which incorporates secondarydata sources from engineering, computer science, and market research domains, the study addresses thefollowing questions:1. What are the GHG emissions attributable to device, network, and datacenter infrastructure inservicing end-uses consumed on ICT devices, in a US consumer context in 2012 and 2017?2. How do network and datacenter emissions compare to emissions due to using devices in thehome?3. Which end-uses are the largest contributors to emissions, and how is this likely to change lookingforward to 2017?The studies collectively argue for a reorientation towards higher-level integrated analyses, as op-posed to point estimates of individual products or behaviors, in order to sidestep issues of study incom-patibility, to allow for better flexibility in adapting to changing market conditions, and to open routesinto connecting fundamental LCA results with research questions regarding higher-order environmentalimpacts of ICT. The significance of each of these studies is summarized in the following section.1.2 SignificanceThe research in this dissertation makes several original, significant contributions to knowledge, outlinedas follows. First, the meta-analysis of previous LCA studies of desktop PCs presented in Chapter 3 is todate the only published quantitative synthesis of LCA studies of ICT devices. It satisfactorily resolves anambiguity in the literature which arose due to large differences in modeling assumptions across differentstudies. Through a thorough review of prior studies of operational energy consumption of desktopPCs, including field measurements, surveys, and data from ENERGY STAR, the study establishes a5reasonable range for operational energy consumption of a desktop PC, and identifies unrealistically lowassumptions for operational energy among two studies which reported production impacts exceedingoperational impacts. The study concludes that energy and emissions due to operation of a desktop PCusually exceed those due to its production. In addition, the study showed that differences in productfrom factor, e.g. between a large workstation-class desktop and a small integrated desktop, could be adominant factor influencing the study result.Second, the study of embodied emissions of consumer electronics in Chapter 4 fills several impor-tant research gaps. First, it is the first peer-reviewed study of a tablet, netbook, compact desktop, andthin-client device, and the first study to include a wide cross-device comparison of these devices along-side more highly studied product categories, namely desktop, laptop, and monitor. It demonstrates thatsmaller form-factor products are less impactful in terms of embodied emissions, and that newer productsare more materially efficient than older products due to higher levels of integration, leading to reductionsin emissions. Second, it makes an important contribution of primary data: all of the study’s laboratorymeasurements and a complete specification of the model were provided in the published study’s sup-porting information under an Open Access license, making the results fully reproducible and availablefor adaptation by other researchers. Third, it analyzes product environmental reports published by Ap-ple [46], and finds cross-device trends that are consistent across both the study’s results and Apple’sreports, which increases confidence in the quality of Apple’s reports, despite their lack of peer reviewor disclosure of modeling assumptions and data sources. Fourth, it proposes a statistical linear model toestimate embodied emissions which suggests a route towards first-order impact estimation that sidestepstime-consuming, expensive LCA.Finally, the study of connected media in Chapter 5 significantly expands the breadth and depth of ourknowledge regarding the impacts of emerging behaviors on emerging platforms. It is the first study tocompare the impacts of broadcast TV, video-on-demand, and other online and offline end-uses, and thefirst to compare impacts of end-uses using TVs against those using PCs, tablets, and smartphones. Thederivation of energy and emissions intensity of fixed and mobile networks and datacenters in 2012 and2017 is by itself a useful contribution to the literature which will improve the ability of other researchersto estimate the impacts of network-enabled end-uses In addition, this study appears to be the first tointegrate behavioral data from market research firms alongside energy and emissions data, which wasthe key factor enabling calculation of total emissions due to each end-use on each platform; the data are6presented transparently and completely to facilitate easy adaptation by other researchers.1.3 Thesis overviewThe thesis is structured as follows. Chapter 2 provides background and context for the research. It in-cludes an overview of current trends in the high-tech sector; a review of conceptual frameworks throughwhich environmental impacts have been discussed; a review of energy accounting and LCA, includingprior application to the estimation of the impacts of ICT products and services, and knowledge gapswhich are addressed in this dissertation; and a discussion of the methodological limitations of LCA,alternative approaches, and constraints on the validity of the findings. Taken together, these topicsprovide the necessary background for situating the research in both the larger debate on LCA method-ology as well as the more specific challenges to environmental impact assessment posed by the rapidtransformations in the ICT sector.Chapter 3 presents a meta-analysis of life cycle assessment studies of desktop PCs. This chapter isan adaptation of a manuscript which was originally published in the Journal of Industrial Ecology in2012 [1], co-authored with Dr. Milind Kandlikar. Supplementary material for this study is in AppendixA.Chapter 4 presents a study of the embodied GHG emissions of 11 electronics products and derivationof a first-order model for estimation of embodied GHG emissions based on simple product characteris-tics. This chapter is an adaptation of a manuscript which was originally published in the journal Envi-ronmental Science and Technology in 2013 [2], co-authored with Dr. Milind Kandlikar. Supplementarymaterial for this study is in Appendix B.Chapter 5 presents a study of the GHG emissions due to consumer ICT end-uses using televisions,PCs, tablets, and smartphones, considering emissions due to device, network, and datacenter. Thischapter is an adaptation of a manuscript, co-authored with Dr. Eric Masanet and Dr. Milind Kandlikar,which is not currently published. Supplementary material for this study is in Appendix C.Finally, Chapter 6 summarizes the conclusions of each study and situates them within the broaderresearch context, including a discussion of methodological limitations, directions for future work, andopportunities for operational application of the methods and findings put forward in this dissertation.7Chapter 2Context and backgroundThis chapter begins with an overview of different conceptual frameworks through which the environ-mental impacts of ICT have been studied. Next, the life cycle assessment methodologies applied in thisdissertation are introduced, including methodological fundamentals related to the study of ICT productsand services, trends in the ICT sector and associated research challenges, a review of prior work, and anoverview of the resarch approach applied in this dissertation. Finally, the methodological limitations ofLCA and of the studies on this dissertation are discussed.2.1 Conceptualizing impacts of ICTThe research in this thesis primarily applies environmental life cycle assessment, a quantitative impactaccounting methodology. This section reviews approaches to understanding environmental impacts andICT, and situates LCA within this context. Two conceptual frameworks are introduced: a three-leveltaxonomy, and an economic decomposition approach, with focus on the latter as it can be quantitativelylinked to the life cycle assessment-based approach used throughout this dissertation.2.1.1 The three-level taxonomyA review by Erdmann and Hilty [47] identifies two prominent categories of conceptual frameworkfor considering the large-scale impacts of ICT systems on the environment: economic frameworksthat examine ICT as a driver of technological efficiency, economic growth, and structural change;and a three-level taxonomy identifying direct/first-order impacts from the life cycle of ICT hardware;indirect/second-order impacts resulting from the application of ICT on other sectors, e.g. smart grids8and smart buildings; and systemic/third-order/feedback impacts, including rebound effects and large-scale societal change, caused by the “emerging effects of ICT in the economic system”, e.g. changesin travel patterns due to telecommuting. This taxonomy was first proposed by Berkhout and Hertin ina 2001 study [48] which categorized many of the ways ICT applications could impact the environment.The original characterization from Berkhout and Hertin is excerpted below, though other studies differsomewhat in their classification of different phenomena:• First order impacts: direct environmental effects of the production and use of ICTs(resource use and pollution related to the production of ICT infrastructure and devices,electricity consumption of ICT hardware, electronic waste disposal)• Second order impacts: indirect environmental impacts related to the effect of ICTs onthe structure of the economy, production processes, products and distribution systems;the main types of positive environmental effects are dematerialization (getting moreoutput for less resource input), virtualization (the substitution of information goods fortangible goods) and ‘demobilization’ (the substitution of communication at a distancefor travel)• Third order impacts: indirect effects on the environment, mainly through the stimu-lation of more consumption and higher economic growth by ICTs (‘rebound effect’),and through impacts on life styles and value systems.The three-level taxonomy has been influential, used in large reports by Forum for the Future [49] andthe European Commission [50], and recently in what Erdmann and Hilty call the “second wave” of ICT-environment studies [47], a series of reports contrasting the relatively small direct impacts of ICT withpotentially large indirect environmental benefits due to the application of ICT [27, 30, 33, 51, 52, 53].Each of these studies considers a set of applications, calculates environmental benefits possible dueto ICT improvements in these applications using LCA or simplified carbon footprinting approaches,and imposes an uptake percentage to predict an overall potential benefit. Each concludes that ICTcould deliver reductions in GHG emissions several times larger than the direct emissions of the ICTsector, though none of the studies incorporate behavioral models to predict actual likely outcomes. TheSmart2020 report is a representative example [30]; the reported potential savings of 30% in electricitytransmission losses due to the introduction of smart grids is not a model output, but rather an assumption.Erdmann and Hilty present the only quantitative treatment of rebound effects in a model that usesthe three-level taxonomy [47]; they used a system dynamics model that attempts to account for feedbackbetween efficiency gains and demand for specific services, finding a mixture of positive and negativeconsequences for greenhouse gas emissions, and point towards the adoption of economic modeling9techniques to account for direct, indirect, and macroeconomic rebound effects. Aside from this study,applications of the three-level taxonomy appear to have been largely conjectural or aspirational. Theunderlying argument, e.g. that ICT can unlock large savings on the order of 15 to 20% of globalGHG emissions by 2020 [30] – is certainly important given the magnitude of the potential savings.However, with the potential efficiency gains spread across multiple domains – power grids, motors,logistics, and building operations, in this case – recommendations for achieving these gains are bynecessity high-level. Efficiency gains, the major potential emissions wedge offered by ICT, can beunderstood through economic frameworks which offer more precise terminology and tools. Behaviorchange, which underlies the higher-order impacts theorized to be large in many studies that use the three-level taxonomy [54], can be modeled through economic frameworks as well. Accordingly, economicframeworks are discussed below.2.1.2 Economic frameworksChanges in environmental impacts may be understood through a resource economics framework, suchas was applied in the Digital Europe study of the dematerialization potential of ICT [55], which exploredthe macroeconomic influence of the ICT sector on the broader economy in Germany from 1991 to 2000.That study’s approach, based on a tradition of economic disaggregation approaches from which the IPATand related equations developed [56], is summarized as follows. Suppose an economy is divided inton sectors each having economic output Yi, such that the GDP of the economy Y is equal to the sum ofeach sector’s output:Y =n∑i=1Yi (2.1)Consider an environmental impact, such as overall GHG emissions, or use of energy, or other re-source, which is specified by I, which is likewise attributable to the impacts of the same n sectors:I =n∑i=1Ii (2.2)The overall impact, I, can be decomposed as follows:I = Yn∑i=1YiY·IiYi(2.3)10Here Y is again the overall GDP of the economy; Yi/Y is the share of economic output attributableto sector i; and Ii/Yi, the impact per unit of output, is the resource intensity of sector i. We may definetechnology level, T , to be the inverse of resource intensity, so that an improvement in technology levelcauses a decline in impact per unit of economic output, and likewise define Ti to be the technology ofsector i.Possible impact trajectories are illustrated in Figure 2.1, which may represent an entire economy,or just an economic sector. At a given level of consumption (equivalently, economic output), Y , theimpact is constrained to pass through the diagonal line imposed by the current level of technology, T .Thus, changes in consumption alone, known as the scale effect (either growth or contraction) causeimpact to rise or fall along the line specified by T . Changes in technology with consumption heldconstant will push the impact straight up or down. Most commonly, consumption and technology changesimultaneously; growth in consumption accompanied by technology improvements is known as strongde-coupling, if net impact declines, and weak de-coupling, if net impact grows.ConsumptionImpactYITStrong de-couplingWeak de-couplingScale eectDecreased intensityGrowthContractionIncreased intensity Re-couplingTechnologyFigure 2.1: Illustration of impact trajectories under consumption Y and technology T .The Digital Europe study considered two ways in which ICT could impact the environment: through11structural change in which the lower-impact ICT sector gains a larger share of the economy, and throughinduced technological change in other sectors [55]. The latter effect is an equivalent though moreprecise framing of the “second-order” efficiency gains in the three-level hierarchy discussed above; forexample, the 30% decline in grid emissions postulated in the Smart2020 report could be expressed as animprovement in technology in the energy sector which causes the total impact of that sector to decline by30%. However, a meaningful expression of the impacts of technological change in an economic sectormust also characterize the expected change in consumption, whether through rebound effects, or simpleeconomic growth. Such effects can be murky; despite contemporaneous growth in ICT investment andimprovements in resource productivity in non-ICT sectors, the Digital Europe study was not able toshow a conclusive statistical link between the two [55].Macroeconomic impacts of ICT have been studied using economic frameworks similar to the oneshown above, especially in the context of productivity [57, 58, 59], changes in economy-wide energyintensity due to ICT [60, 61], and the concept of the information society [62, 63, 64], but are beyondthe scope of the research in this dissertation. Likewise, large non-linear changes to society (some ofthe “third-order effects” in the three-level hierarchy) are not easily captured with simple linear modelsand are similarly out of scope. However, economic decomposition provide a useful framework for ofcategorizing and comparing impacts of ICT products and services in aggregate, which can be illustratedthrough the history of study of the effects of ICT on travel.The potential for telecommunications to reduce the need for travel has been studied since at leastthe early 1980s, and arguably even since the invention of the telegraph [65]. Salomon reviewed thetopic in 1986 [66], and identified a complex relationship including stimulative effects of telecommuni-cations on travel which counteract substitution effects. Telecommuting and teleconferencing have clearenvironmental advantages over physical travel when compared on a functional unit basis, i.e. work-ing from home does in most cases offer energy and GHG emissions advantages over commuting to anoffice [67, 68], and the replacement of a physical trip with videoconferencing does result in substan-tial emissions savings [44]. However, empirical analysis by Mokhtarian showed that despite increasedavailability of ICT services, most indicators of travel demand continue to climb [69]; furthermore,there is compelling evidence of a complementary or even stimulative relationship between ICT andtravel [70, 71]; use of ICT may indeed diminish the need for some trips, but simultaneously generatesdesire for other trips.12Consider Equation 2.3, but simplified to consider only the impact of a specific sector, Ii:Ii = Yi ·IiYi(2.4)This formulation, though very simple, isolates the critical distinction between efficiency or techno-logical improvements, which affect IiYi , and overall consumption, which affects Yi. A more pragmaticversion of this equation can be applied to specific services, such passenger travel, measured in vehicle-kilometers (vkm). In that case, considering impact in terms of GHG emissions:Itravel[kg CO2e] = Ytravel[vkm] ·ItravelYtravel[kg CO2evkm](2.5)The impact due to all travel is equal to the total amount of travel consumed, Ytravel, measured in vkm,multiplied by the average fleet GHG intensity, Itravel/Ytravel, measured in kg CO2e/vkm. The latter, beingexpressed in terms of emissions per functional unit, can be determined using an LCA approach, andwould be influenced through technological efficiency gains, such as the introduction of lower-emissionsvehicles. The former is driven entirely by human behavior and must be studied empirically, perhapsthrough the use of equilibrium economic models, surveys, or field measurements. In the case of theinfluence of ICT on travel, while telecommuting was shown to be a more efficient activity, such that aunit decline in travel accompanied by a unit increase in telecommuting would lead to a net reduction inemissions, growth in total travel demand caused a continuing rise in overall emissions.Some studies consider the impact of a product or service per functional unit, usually performedusing life cycle assessment, while others estimate the level of consumption of a product or service,derived empirically or through models of human behavior. Studies of overall impact must combine thetwo. For example, when considering the impacts due a product such as a desktop computer, the overallimpact is the product of the installed base of the product, IB, and the unit impact per product:Idevices[kg CO2e] = IB[devices] ·Idevices[kg CO2e]IB[devices](2.6)= IB[devices] · Idevice[kg CO2edevice](2.7)13A new quantity, Idevice, is defined to be the unit impact per device which can be determined usingLCA, but the installed base IB is an empirical quantity which may be estimated using sales data andother empirical sources.Likewise, when considering the impacts of network-enabled ICT services that involve communica-tion, the network portion of the impact may be modeled as follows, where data traffic associated withcommunication is measured in gigabytes (GB):Icommunication[kg CO2e] = Traffic[GB] ·Icommunication[kg CO2e]Traffic[GB](2.8)= Traffic[GB] ·GI[kg CO2eGB](2.9)Again a new quantity is defined, GI, to represent the GHG intensity of communication, which canbe multiplied by the total amount of traffic to obtain an estimate of total impact. The energy intensity ofthe internet, in kg CO2e/GB or related units, is a widely used statistic (see review in [72]) and a centralfeature of the study in Chapter 5 of this dissertation.Life cycle assessment is a methodology for estimating impacts per functional unit of products orsystems (e.g. Idevice or GI above). The following sections introduce LCA and its application to the studyof ICT products and services.2.2 LCA and ICT productsThis section introduces the energy accounting and life cycle assessment methodologies used in thisdissertation to quantify impacts of ICT products, discusses recent trends in the ICT sector and theirresearch implications, and motivates the research in Chs. 3 and 4 of this dissertation. ICT services arediscussed in Section Methodological approachesThe unit impact quantity, derived in Equation 2.7, represents the total impact of a device over its lifecycle. This can be conceptualized by considering a set of life cycle phases, P, which may be broadlydefined as ‘production’, ‘operation’, and ‘disposal’. The production phase may itself include severalsub-phases, such as raw materials extraction, parts manufacture, final assembly, and transportation to14consumer. The impact of the device is the sum of the impacts due to each life cycle phase:Idevice = Idevice,production + Idevice,operation + Idevice,disposal (2.10)The basic premise of LCA is that impacts of a system can be determined by specifying that system interms of physical flows and applying characterization factors which linearly map those flows to impacts,and while there are some issues with this premise which are discussed in Section 2.4, LCA is a widelyused and accepted method for quantifying impacts in many research domains. Consider the impact ofa device, D, which can be described as a number of flows of substances f , such as steel, electricity,semiconductor chips, etc., where each flow has quantity Q f . In LCA, the impact of the device ID isdetermined as follows:ID[impact units] = ∑f∈DQ f [flow units] ·CF f[impact unitsflow units](2.11)The collection of flow units ∑ f∈D Q f is called the life cycle inventory (LCI) of D. The character-ization factors are collectively defined through one of many possible life cycle impact assessment(LCIA) schemes, which include IPCC [73], TRACI [74], Impact 2002+ [75] and Impact World+ [76],EDIP [77], and others; these schemes each allow for calculation of impacts in a number of dimensionswith different impact units, including global warming potential (in kg CO2e), ozone depletion potential(in kg CFC−11e), eutrophication potential (in kg NOxe), and many others. LCA modeling is hierarchi-cal; complex non-physical flows, such as energy (in kWh) or travel (in vkm), may be defined in termsof fundamental physical flows. Matrix formulations are required to determine the impacts of complexsystems and especially to account for circular flows; details are available in textbooks [78].The conventional form of LCA discussed above is known as ‘process-sum’ or simply ‘process’LCA [79], which is a bottom-up method, so called because the total estimate of impact is obtainedby summing the impacts of components. Bottom-up methods offer high degrees of precision, but arevulnerable to truncation errors due to components excluded from the modeling boundary, which arepotentially substantial [80, 81]. Section 2.4 discusses this issue and alternative approaches. Impactquantification in practice uses scientific software and databases to define characterization factors, and tospecify models which describe finished materials and components (e.g. steel, electricity, etc.) in terms15of more fundamental industrial processes and physical flows to environment. Ongoing fundamental re-search in the LCA field aims to improve characterization factors, industrial models, and environmentaldamage models. The research challenges in determining the impacts of a product using LCA lie in ac-curately describing the components of the product, and in obtaining and applying appropriate industrialmodels which model those components in terms of physical flows. Operational energy consumption isdiscussed below to illustrate the mechanics of an LCA calculation.Operational impacts of electronics are determined by the amount of energy they consume, alongwith the characteristics of the energy source, typically a power grid. Applying the framework above,operational impacts are assessed by defining a quantity Q of electricity and multiplying by a characteri-zation factor. For energy-using devices, the quantity of electricity consumed in one year under averageconditions is often called the unit energy consumption (UEC), measured in kWh/yr [4, 5, 82]. Thecharacterization factor which converts an electricity flow to an impact quantity is known as the gridemissions factor, EF. Operational impacts are thus expressed as follows:Idevice,operation[kg CO2eyr]= UEC[kWhUSAyr]·EF[kg CO2ekWhUSA](2.12)Here the model requires an additional assumption: the location of the power grid from which theenergy was drawn, e.g. USA in the equation above. Grid emissions factors are a specific example of acharacterization factor, which converts a physical flow, in this case one kWh of electricity generated inUSA, with an impact, in this case global warming potential, measured in kg CO2e, obtained through theuse of empirically-driven models that describe the current physical energy infrastructure in the USA.Unit energy consumption, a required input for calculating operational impacts in a life cycle assess-ment of an energy-using product, is of key importance and is thus discussed in detail below.Device unit energy consumptionThe unit energy consumption is a major determinant of device overall impact, and is thus an importantcomponent of a device LCA study, but it may also be studied on its own for studies which considerenergy use only. In its most basic formulation, the energy consumed by an energy-using device over aperiod of time, defined by times t1 and t2, is equal the integral of the instantaneous power consumed by16the device over that interval:E =∫ t2tiP(t)dt (2.13)Power, P(t), may fluctuate over time. However, if the average power draw over a period of timeis known, then energy may be obtained by multiplying average power draw by a time amount. In thiscase, energy per year can be obtained by multiplying average power draw over the year by the numberof hours in a year:UEC[kWhyr]= Pavg [kW] ·8766[hyr](2.14)In most cases, device average power over a year will not be known. However, researchers haverecognized that while device power draw will fluctuate over time, most devices have steady operatingpower draws in one of several power modes, such as ‘active’, when the device is being actively usedat its highest level of functionality; ‘idle’, when the device is on but not being actively used; ‘sleep’,when the device is on but in a low-power sleep mode; and ‘off’, when the device is switched off, butstill plugged in (again, for examples see [4, 5, 82]). Total energy consumption is then obtained througha sum of the energy consumed in each power mode:UEC[kWhyr]= ∑i∈PMEi = ∑i∈PMPi [kW] · ti[hyr](2.15)Here Pi represents the average power draw while in power mode i ∈ PM, where PM representsthe set of power modes; power draws can be obtained through laboratory measurement of equipmentunder simulated operational conditions using power meters, and are sometimes reported by productmanufacturers. The time spent per year in each power mode is ti; this is a strictly empirical quantitywhich depends on human behavior, and on device power management features; for example, manydevices will automatically enter a sleep mode after being left idle, but this may depend on whetherpower management features are enabled. Time spent can be estimated via consumer surveys or fieldmeasurements.Koomey applied a similar approach to his calculations of energy due to servers in data centers, butsince servers are usually on all the time, each server’s energy was calculated using an average powerestimate as in Equation 2.14 [23, 83]. In addition, the study of data centers in particular often requiresestimates of the often significant overhead energy consumed due to cooling, power distribution, and17other non-computational energy use. A common formulation is as follows:Etotal = Eactive +Eoverhead (2.16)= (Eactive +Eoverhead) ·EactiveEactive(2.17)=(Eactive +EoverheadEactive)·Eactive (2.18)= PUE ·Eactive (2.19)Here Eactive represents energy consumed due to operation of service, storage, networking equipment,and other devices used in the processing of information; Eoverhead is all other operational energy con-sumed in the data center; and PUE, the power usage effectiveness, is the ratio of total energy to activeenergy. PUE has emerged as a rule of thumb for assessing the efficiency of data centers, where PUE of1.0 is the theoretical no-overhead minimum; it can also be used to facilitate first-order estimates whereonly active energy is known. Koomey applied a US and global PUE of 2.0 in his 2007 study [83]; for2010, he estimated US average PUE to have declined to between 1.92 and 1.83 [23], based on third-party surveys of US data centers. Highly efficient enterprise or cloud-scale data centers can have PUEswell below the average; Google reports an average PUE of 1.12 over 2013 for their datacenters, havingdeclined from 1.2 in 2009 [84]; Facebook reports PUEs of 1.09 over the last year for each of its Oregonand North Carolina data centers [85].In summary, the quantification of impacts of ICT products relies on several pieces: first, an inventorydescribing the material and energy inputs which arise in various phases of the product life cycle; second,a database or dataset of industrial models which allows for expression of the material and energy inputsin terms of fundamental physical flows to environment; and third, a life cycle impact assessment schemewhich converts these flows to a relevant impact category through a set of characterization factors. Inpractice, the latter two steps may be satisfied using scientific databases such as ecoinvent or GaBi alongwith LCA software which implements characterization schemes, though there is an ongoing need toimprove industrial models of ICT production and disposal processes. If these are available, then themajor research challenge is in obtaining an accurate model of the product’s components and inputs, aswell as the behavioral characteristics of the product’s users which influence energy consumption. Due18to rapid change in the ICT sector, such models and data sources may rapidly fall out of date, as newproducts are introduced and new behaviors emerge. The following section briefly discusses some driversof change in the ICT sector and their research implications.2.2.2 Industry profile: On Moore’s Law and energyICT is a dynamic, fast-moving field, characterized by short product lifespans and frequent emergenceof new products. Regular technological improvements in semiconductor fabrication processes are themajor driver of change in the electronics sector. This process is known as ‘scaling’, as each improve-ment enables a higher area density of transistors on silicon die, thus allowing circuit geometries to bescaled downward in size. The quantity of transistors per unit area has increased exponentially, doublingapproximately every two years, following an empirical trend known as Moore’s Law, originally formu-lated in 1965 [86]. The transistor is a fundamental building block of both computation and memory, so achip with more transistors has more computational functionality. Scaling has reduced the cost per com-putational instruction by a factor of 100 per decade for the last forty years [87], meaning that modernelectronics are significantly less expensive and significantly more powerful than older counterparts. Ac-cording to an industry report, biennial semiconductor fabrication process upgrades will continue throughat least 2028 [88].When transistors are fabricated at smaller sizes, their operational power consumption declines inproportion to their area (to first order), according to Dennard’s scaling law which describes the physicsof scaling MOSFET transistors [89]. Thus, while transistor density and computational power per devicehave grown exponentially over time due to scaling, electrical power density has remained roughly con-stant, implying that computational power per unit of operational energy has also grown exponentially.Koomey explored this trend empirically, and showed that the number of computations per operationalkWh in computers has been doubling about every 1.6 years since the dawn of the computing era in the1940s [90].Semiconductor scaling has enabled an unrelenting march of improvements in consumer and busi-ness electronics products which offer greatly enhanced functionality relative to previous generations,roughly following the 2-year cycle of semiconductor fabrication process improvements. Electronicsproducers offer products in various form factors to target different use cases, exploiting a tradeoff be-tween higher-performance and higher-energy-consuming components typically used in desktop PCs,19and lower-performance but lower-energy-consuming components typically used in mobile electronics.The smartphone emerged as as a computing platform in the mid-2000s; its processor was sufficientlypowerful to allow the device to offer levels of performance sufficient to support web and media appli-cations that were formerly the exclusive domain of PCs. The tablet has emerged as another computingand media platform bridging laptops and smartphones in the product-space; soon after the “phablet”emerged in the space between smartphones and tablets, as did hybrid devices which have characteris-tics of both tablets and laptops. Additionally, within a product category such as laptop or desktop, onecan find a range of devices which may have significantly different sizes, costs, performance, and en-ergy characteristics [5]. This proliferation of devices creates challenges for researchers, as the existingenergy and LCA literature has limited coverage of newer electronics products.2.2.3 Knowledge gaps and research approachReviews of prior applications of LCA to ICT devices were conducted in 2011 [19] and more compre-hensively in 2014 [91]; the literature reviewed in the latter includes coverage of PCs, monitors, phones,servers and datacenters, semiconductors, network devices, and TVs. Studies of ICT services and end-uses, which were also included in the 2014 review, are discussed in Section 2.3.LCA studies may vary on several dimensions, as cataloged in [91]: which type of LCA methodologyto use; how to define the scope and boundary of the devices or systems under study; how to model thecontents of the device or system, whether through primary disassembly, proprietary manufacturing data,or other data sources; what data source to use to model upstream processes of manufactured componentsand inputs, such as an LCA database like ecoinvent or GaBi, or proprietary industrial data sources; whatimpacts to calculate and what set of impact characterization factors to use; and naturally what high-levelgoals underpin the study. This multi-dimensional research space has many frontiers. The exploration ofhybrid LCA methods to calculate device impacts while correcting for truncation error are an importantmethodological advance, having been developed by Eric Williams and others through case studies of adesktop PC [13] and laptop PC [92]; see discussion in Section 2.4.2. Industrial data describing electroniccomponent manufacturing processes underpins all LCA studies and is in need of continual refreshment.Better models of recycling and disposal processes are needed, especially informal recycling activitieswhich are not currently modeled in any LCA frameworks. In addition, empirical data relating to deviceuse and lifespan, e.g. through surveys and field measurement campaigns, is needed to accurately model20device operational impacts.The research in this dissertation addresses the challenge of interpreting this literature in an opera-tional context. Due in part to these multiple dimensions of variation, different LCA studies of similarproducts may produce significantly different results [93], producing challenges for ICT-using organiza-tions and individuals who rely on such studies to support decision-making. In the case of desktop PCs,which are the most-studied device [91], numerical results of existing LCA studies estimating green-house gas emissions and primary energy consumption vary by more than an order of magnitude, andsome studies show dominant impacts from the operational phase while others show dominant impactsfrom the production phase. Chapter 3 conducts a meta-analysis of 13 such studies to examine sourcesof variation, identify the source of the disagreement regarding the relative impacts of the productionand operational phases, and produce reasonable bounds on the greenhouse gas emissions and primaryenergy demand of a typical desktop PC used under average operating conditions. Due in part to the chal-lenge in identifying modeling assumptions and input data sources, this study remains the only publishedquantitative synthesis of LCA studies of ICT devices.The study in Chapter 4 of this dissertation addresses the additional problems caused by deviceproliferation. Despite the number of LCA studies – about 60 reviewed in [91] – it is unfortunatelydifficult to draw general conclusions regarding the impacts of ICT devices from this literature, preciselybecause there are so many dimensions of variation. During the design phase of the ICT systems for theCentre for Interactive Research on Sustainability (CIRS) building at UBC in 2010, stakeholders inquiredabout the relative impacts of available products in order to help guide decisions towards lower-impactoptions. Product coverage in the literature was an immediate concern; some of the products underconsideration, such as tablet PCs, lightweight desktops, and thin client devices, had not yet been thesubject of an LCA study. Comparison was another; understanding the impacts of provisioning laptopsinstead of desktops (for example) requires a scientifically valid comparison of the two with consistentdata sources and modeling assumptions. At the time, only the ecoinvent V2 database, which includesfully transparent studies of a desktop PC, laptop PC, LCD monitor, keyboard, mouse, and networkdevice [94, 95], provided a consistent modeling framework allowing for comparisons of one deviceagainst another; this was adapted for a screening study of the CIRS ICT systems [96] which motivatedthe study in Chapter 4. This work includes estimates of the embodied GHG emissions and primaryenergy of 11 ICT devices with device data from primary teardowns and component data and modeling21assumptions from ecoinvent. The framework allows for comparisons of the studied devices against oneanother and also against the older-generation ICT devices modeled in ecoinvent. In addition, Apple’sproduct carbon footprints are compared against the study’s results.Research into impacts of ICT products lays the foundation upon which we may assess the broaderimpacts of ICT usage in society. In many cases, it is desirable to focus the analysis not on products, buton the functions for which the products are used, which implies a different and possibly significantlylarger modeling scope and a different set of methodological challenges, discussed in the next section.2.3 LCA and ICT servicesServices and end-uses may also be assessed using a life cycle assessment approach. An end-use is de-fined to be a specific behavior for which ICT devices are used, such as reading the news or watchingvideo, while an ICT service is defined to be a collection of software and infrastructure which users canaccess, such as Netflix. In other words, a service can be a means through which an end-use is per-formed. Some end-uses like using productivity software on a computer are relatively simple to model,as they involve only a portion of the operational impacts of the computer, but many end-uses involve thetransfer of data over communication networks, and thus require some means of assessing the impacts ofthose networks and allocating an appropriate share. Such models invariably have larger scopes than theproduct LCAs discussed in the previous section, because they must at minimum account for the impactsof the devices through which the services are consumed, as well as all of the infrastructure systems thatsupport the services. Methodological approaches to assessing the impacts of ICT services are discussedbelow, followed by an overview of domain-specific trends and research challenges, and a discussion ofprior work and the research approach taken in Ch. Methodological approachesImpact of a service can be modeled as follows:Iservice[impact units] = ∑i∈devices,systemsIi,service (2.20)Each device and system which is involved in the delivery of the service must be included in the model.In each case, an impact of the device or system which is attributable to the service, Ii,service, is defined.22Suppose the impacts of the device or system may be allocated linearly according to some functionalflow with quantity Q, which could be time spent, or data traffic generated, or a similar metric. Thenimpact of the service can be expressed as follows:Iservice[impact units] = ∑i∈devices,systemsIi[impact units] ·Qi,service[flow units i]Qi,total[flow units i](2.21)For each device and system i within the study boundary, the overall impact Ii is required, as well as thetotal quantity of functional flow Qi,total, and the quantity of functional flow attributable to the serviceQi,service. The fraction Qi,service/Qi,total is a ratio beween zero and one which expresses the portion of thedevice or system attributable to the service. Alternatively, the equation may be re-stated as follows:Iservice[impact units] = ∑i∈devices,systemsIiQi,total[impactflow units i]·Qi,service[flow units i] (2.22)The data requirements are unchanged, but it is sometimes more convenient to calculate an impact in-tensity per functional flow for the device or system, Ii/Qi,total, and then multiply this intensity by theamount of functional flow attributable to the service, Qi,service. This formulation is equivalent to the ex-ample in Equation 2.8, which expressed impacts in terms of GHG emissions, functional flow in terms ofGB of data transfer, and defined GI, the GHG intensity, to be equivalent to total GHG emissions dividedby total data transfer.In some cases, it is useful to differentiate between the impacts of one usage of the service, and theaggregate impacts of many uses of the service. In order to facilitate this comparison, assume that Iservicerepresents the total aggregate impact from all uses of the service, and likewise Qi,total is the aggregatefunctional flow on device or system i due to all uses of the service. The service may be defined in termsof a functional unit, such as an hour of video, or one web page, and so on. Let the total consumption ofthe service be Yservice, measured in service functional units. In that case, the overall impact due to theservice can be expressed as follows:Iservice[impact units] = Yservice[service units] ·IserviceYservice[impactservice unit](2.23)In other words, a new quantity I/Y is the impact per functional unit, where the aggregate impact Iis the product of Y and I/Y . Impact per functional unit may be restated as follows:23IserviceYservice[impactservice unit]= ∑i∈devices,systemsIiQi,total[impactflow units i]·Qi,serviceYservice[flow units iservice unit](2.24)This formulation makes a distinction between the functional unit of the service being considered,and the functional unit of any devices and systems that the service requires. Consider, as an example, anetwork service that accesses the internet. Where system i is the internet, suppose internet impacts maybe allocated according to functional flow of data, so that Qi is measured in GB. The impact intensity ofthe internet, Ii/Qi,total, may be known from third-party studies. Suppose the network service is consumedby the hour, so that Y measures the number of hours. Then, Qi,service/Yservice is a ratio that maps thefunctional unit of the service to the functional unit of the internet and in this case is measured in GB/hr.Iservice/Yservice measures the impact of the service per hour; and Iservice is the aggregate impact of allhours of the service consumed.This framing is simple but highlights an important distinction between the impact per functionalunit, I/Y , and overall impact, I. When comparing the impacts of two different services, it is possiblefor one service to be both more efficient (i.e. I/Y is lower) and also more impactful in aggregate (i.e. Iis larger), if the differences in consumption Y are sufficient to overwhelm efficiency gains. An exampleof this is the rebound effect, well-known in the study of energy efficiency, in which the lower cost of aservice allow people to consume it more.Assessments of specific ICT services tend to be integrated models, incorporating prior estimatesof impacts for devices, networks, and datacenters. The field is characterized by the usage of energyintensity of data transfer as a means of allocating impacts due to networks, including the Internet; thisapproach was first applied by Koomey in 2004 [97]; see a 2014 review by Coroama and Hilty for moredetail [72]. The determination of allocation shares for devices and systems that a service uses is aprimary research challenge in the assessment of impacts of ICT services, exacerbated by the speed atwhich new ICT services arise and ICT-users’ behavior changes. The following section discusses currenttrends related to ICT end-uses and associated research challenges.242.3.2 Industry profile: connected media and the cloudThree key trends are significantly changing consumer media consumption habits. First, TV sets arebecoming connected to internet services, either through game consoles or connected set-top-boxes, orthrough integrated functionality within products known as connected TVs or smart TVs [98]. Theportion of global TVs connected to the internet was estimated to be 12% in 2013, forecast to grow to 27%in 2018 [99]. The internet link allows for streaming of video content over the internet, either throughvideo-on-demand services which are typically run by telecommunication network operators such ascable TV subscription providers, or through internet streaming services like Netflix or YouTube. Highuptake of these services along with high bandwidth of video relative to other forms of internet trafficgenerates disruptive effects on internet traffic patterns, and will certainly drive future infrastructuredevelopment, especially in provisioning high-speed internet access to the home [100]. Laptop anddesktop PCs are also major platforms for the consumption of streaming video via internet.Second, tablets and smartphones are both gaining prominence as media platforms as well [98, 101].These devices rely on either Wi-Fi or mobile (cellular) network connections. Explosive growth in datatraffic due to these devices on both Wi-Fi and mobile networks is forecast in coming years, driven bydemand for high-bandwidth internet services, but also enabled by infrastructural upgrades, such as theintroduction of 4G or LTE mobile networks which support much higher data rates than previous 3Gmobile networks [102, 103, 104, 105].Third, the emergence of scalable internet platforms such as Amazon Web Services (AWS) alongwith high penetration of broadband connectivity via fixed and mobile links has enabled a proliferationof internet services known as “cloud” services [106]. The National Institutes of Standards and Tech-nology defines cloud computing as “a model for enabling ubiquitous, convenient, on-demand networkaccess to a shared pool of configurable computing resources (e.g., networks, servers, storage, applica-tions, and services) that can be rapidly provisioned and released with minimal management effort orservice provider interaction” [107]. The definition therefore includes any from of software, platform, orinfrastructure which is hosted and operated on a scalable remote compute grid rather than on dedicatedservers and/or client computers, and encompasses a broad range of modern ICT services including so-cial networking (Facebook, Twitter), managed file storage (Dropbox), streaming video (Netflix, Hulu,Youtube), streaming audio (Pandora, Spotify, Grooveshark), productivity software (Google Documents,25Microsoft Office 365), and a wide range of commercial software applications. Cisco estimates that by2017, nearly two-thirds of all global data center workloads will be processed in the cloud, and globalcloud IP traffic will account for more than two-thirds of total data center traffic [108].These combined trends lead to several major developments in the short term:1. Proliferation of connected media devices that access cloud services, including smart glasses andsmart watches, along with increased adoption of tablets, smartphones, and similar devices [109,110, 111, 112].2. Technological convergence in the digital media space, especially with respect to television broad-cast systems which will increasingly merge with internet content delivery systems [100, 113].3. Increased uptake of cloud services that may be accessed from multiple different devices, in whichprocessing and storage infrastructure exists via remote providers [114, 115].These trends create a research need for forward-looking studies that can capture the complexity ofmodern services, which are hosted in flexible compute grids and accessible from multiple devices.2.3.3 Knowledge gaps and research approachThe study of environmental impacts of ICT end-uses began in the early 2000s; prior work is reviewed inTable 2.1. A majority of studies compare an ICT end-use against a physical counterpart; print and digitalmedia is the most common comparison, appearing in eight studies, additionally reviewed in [116].Meetings via ICT were compared against physical meetings in four studies; three studies compared theimpacts of shipping digital media (CDs and DVDs) via e-commerce networks against other distributionmeans. In addition, a more recent trend is the comparison of ICT end-uses against other ICT end-uses,including studies of cloud software and digital media consumption in the home. The study in Chapter 5of this dissertation, which is also listed in Table 2.1 for comparison, falls into this latter category.26Study End-use Platforms Location / data year Functional units Impacts Largest impact[117] Phone system UMTS, GSM Switzerland, 2004 Gbit; person-yr GWP, PED, EI GSM > UMTS[118] Meetings Teleconferencing, travel USA, 2003? Meeting Many Physical > digital[119] Meetings Videoconferencing, travel Japan, 2007? Meeting GWP Physical > digital[44] Meetings Videoconferencing, travel N/S, 2010 Meeting GWP, PED Physical > digital[120] Meetings Videoconferencing, travel Sweden, 2012? Hr; meeting; company-yr GWP, PED Physical > digital[38] Books Print, e-reader Sweden, 2009 Book Many Variable[121] Textbooks Print, e-reader USA, 2002? 40 textbooks GWP, ODP, AP Paper > digital[122] Textbooks Print, online (desktop) Sweden, 2009? 5000 students / 5 yrs GWP Digital physical[123] Magazines Tablet Sweden, 2010 Copy; hour GWP Few readers > many[124] Scholarly journals Print, online (desktop) USA, 2001-10 Article PED Variable[125] E-mail Desktop, laptop France, 2012? 1 Mb Many Numerical result only[126] News Print, TV, PC Switzerland, 2001 News item; person-day EI Paper > online > TV[118] News Print, PDA USA, 2003? Person-yr Many Physical > digital[127] News Print, tablet Sweden, 2007? Reader-yr Many Variable[41] News (video, text online) Desktop, laptop, tablet, smart-phoneUK, 2013? Bit; 10 minutes Energy See Ch. 5[128] Advertising (online) Internet USA, 2006 Impression; US total GWP, energy Numerical result only[42] Business software Local data center, cloud USA, 2013? US business total PED Local > cloud[129] Business software Laptop, tablet; local, cloud N/S, 2013? Typical activity GWP Variable[39] Music (CD, MP3) Physical retail, shipping,download, CD-RUSA, 2008 Album GWP, PED Physical > digital[130] Video (DVD) Physical retail, shipping USA, 2006? 3 movie rentals Many Retail > shipping[40] Video (DVD, streaming) Shipping, PC USA, 2009 Movie GWP, PED Variable[45] Video (Broadcast TV, inter-net VoD)TV, desktop, laptop UK, 2009 Viewer-hr; UK total GWP Broadcast > VoDCh. 5 Video (Broadcast TV, in-ternet VoD, internet video),other internet, offline end-usesTV, desktop, laptop, tablet,smartphoneUS residential, 2012and 2017Device-yr; US consumer to-talGWP, energy See Ch. 5GWP: global warming potential; PED: primary energy demand; ODP: ozone depletion potential; AP: acidification potential; EI: Eco-Indicator pointsN/S: not specified; ?: data year not specified, assumed to be year of publicationVoD: Video-on-demandTable 2.1: Review of studies of ICT end-uses27As with device LCAs, there are many dimensions of variability among these studies, such as choiceof end-use and platform, temporal and geographic scope, definition of functional unit, choice of impactdimension and methodology, and others. The majority of studies focus on global warming potential,energy consumption, or primary energy demand as impact dimensions. Unfortunately, due to rapidchange in the industry, studies can have a limited window of relevance. E-commerce distribution sys-tems for CDs and DVDs, though a relatively recent development, are in decline in favour of streamingservices [131]. Assumptions regarding the impacts due to ICT devices required to pursue and end-useare likewise anchored in a specific time point and may become obsolete; for example, the average im-pacts of using a PC will have declined significantly since 2000 as the proportion of laptops relative todesktops has grown [5]. Likewise, the rapid rise of smartphones and tablets as media platforms threatensthe relevance of studies of end-uses which do not include them. In addition, major categories of ICTend-use have no current coverage in the literature, such as social networking and gaming, while otherssuch as broadcast video and video-on-demand have been studied in limited contexts only.A majority of the studies in Table 2.1 measure impacts on the functional unit basis, e.g. per book,per article, per hour of use, etc.; only a few include some assessment of overall consumption level inorder to assess impacts at an aggregate level. In the context of Equation 2.24, a functional unit as-sessment measures I/Y , while an aggregate assessment measures I at some level of aggregation. Theimportance of aggregation is illustrated by two studies of textbooks compared against digital alterna-tives. The first [121] considered a college context in which students used either an e-reader or used 40paper textbooks during the course of their degree. The second [122] considered a school system with5000 students over five years, using either paper books or desktop PCs. The first study found modestenvironmental benefits to using the e-reader, while the second study found ten to thirty times higher im-pacts from the desktop PC solution. Impacts from desktop PCs are much higher than e-readers, whichin part explains the discrepancy, but in addition, the latter study assumed very high levels of textbookre-use, characteristic of textbooks in a school system, while the former assumed much lower levels ofre-use characteristic of college textbooks. In each case, if this research were to support a decision asto which system to adopt, making the correct choice would only be possible by considering levels ofconsumption in each system.The study in Ch. 5 of this dissertation considers the impacts of digital media end-uses. One exampleof a study in this domain considered the impacts of different modes of music delivery, including physi-28cal retail, shipping CDs via e-commerce, and digital downloads (optionally with burning to CD-R) [39].The study concludes that downloading one album is less impactful, but one cannot conclude that theusage of downloading services leads to lower impacts than the usage of physical retail, because con-sumption of music via each mode may be different, especially given the advent of piracy and streamingmusic services on the internet which have considerably lowered the barriers to music consumption on-line. Likewise, a group of studies of online news via text and video delivered to smartphones, tablets,desktops, or laptops [41, 132] considered impacts per bit of data and per 10 minutes of content browsingand of the system as a whole; while this analysis is valuable in highlighting which components of thedevice, network, and data center systems contribute the most energy towards delivery of these services,it does not assess which end-use/platform uses more energy in aggregate.The study in Ch. 5 has several complementary aims in order to address several knowledge gaps.First, it compares the impacts of several different categories of end-uses which are defined broadlyin order to capture most uses of ICT equipment in the home, thus ensuring that important end-usesare accounted for. Second, it studies each of these end-uses on TVs, PCs, smartphones, and tablets,which gives it significantly broader coverage than the two closest related studies [41, 45], neither ofwhich considers all four of these platforms. Third, it explicitly considers consumption levels at theUS aggregate level, which makes it possible to rank end-use/platform combinations in terms of theiroverall impact. Fourth, using existing industry forecasts and extrapolations of current trends, it estimatesimpacts forward to 2017, in order to increase the longevity of the study.Collectively, the studies in Chs. 3, 4, and 5 apply life cycle assessment methods in an integratedcontext to improve our understanding of the impacts of ICT products and services for use in operationaldecision-support. Key limitations of the LCA methodology and on the validity of these studies arediscussed below.2.4 Methodological limitations and justificationWhile the usage of LCA to estimate impacts of devices, systems, and behaviors is a common practice,the methodology does have several important limitations which should be taken into account, which arediscussed in this section. In addition, given these limitations, justifications are provided for the use ofLCA in this dissertation and in general, and the consequent context in which claims made through theuse of LCA can be considered valid.292.4.1 LCA database and LCIA impact scheme uncertaintiesLCAs are integrative models and usually require the use of an industrial database to model upstreamprocesses and flows, such as due to raw materials extraction and processing. Models may be verycomplex and rely on extensive empirical data, such as the relative mix of energy generation sources usedin a regional power grid. Each process modeled in an LCA database may involve dozens or hundredsof emissions to the environment in various forms. The collective environmental impact of these flowsis defined in LCA through the use of life cycle impact assessment (LCIA) aggregation schemes, whichdefine characterization factors converting physical flows into damage equivalent units in one of dozensof possible impact categories. Some schemes aggregate in synthetic units, while others aggregate interms of equivalent units of a reference substance. Both the empirical models in LCA databases andthe definition of characterization factors in LCIA schemes are potential sources of error and uncertainty.Each is assessed below.The ecoinvent V2 database implements 37 different LCIA schemes, each of which contains sev-eral impact dimensions. Some dimensions like global warming potential are very commonly assessedin LCA studies of ICT devices, as was shown in Section 2.2.3, while many others are less commonlyreported. Table 2.2 shows the output LCIA calculations from the ecoinvent V2 database for selectedquantities – 1 kg of steel, 1 kWh of electricity in USA, and one desktop computer. Four impact dimen-sions are shown: global warming potential, acidification potential, eutrophication potential, and ozonedepletion potential; these were selected because they are among the more common impact dimensionsin terms of prevalence in LCIA schemes. However, they are not all measured in common units. Forthose indicators that are measured in common units, we should expect the output results to be equal toone another, because they are based on identical physical models from the ecoinvent database. In orderto explore the consistency of categories measures in different units, the ratio of electricity to steel andthe desktop computer to steel is calculated as well.The table shows that the two impact categories calculated in identical units, global warming potentialand ozone depletion potential, are mostly numerically stable across several different LCIA schemes;the exception is the TRACI scheme’s ozone depletion potential calculation for a desktop computer.Other categories give varying results. Naturally the numeric results differ considerably because theyare measured in different units; still, for each of acidification potential and eutrophication potential,30steel,low-alloyed,at plant (kg)electricity,high voltage,at grid, US(kWh)desktopcomputer,withoutscreen, atplantElectricityto steel ratioDesktop tosteel ratioGlobal warming potentialIPCC 100-yr [kg CO2-eq] 1.8 0.76 270 0.43 1.5E+02CML 2001 100-yr [kg CO2-eq] 1.8 0.76 270 0.43 1.5E+02EDIP 2003 100-yr [kg CO2-eq] 1.8 0.76 270 0.43 1.5E+02TRACI [kg CO2-eq] 1.7 0.76 270 0.43 1.6E+02Acidification potentialCML 2001 global [kg SO2-eq] 6.9E-03 5.2E-03 1.8E+00 0.75 3.9E+04EDIP 2003 [m2] 1.1E-01 9.4E-02 3.0E+01 0.83 2.4E+03TRACI [moles H+-eq] 3.6E-01 2.7E-01 9.3E+01 0.74 7.4E+02Eutrophication potentialCML 2001 global [kg PO4-eq] 3.9E-03 1.4E-03 2.4E+00 0.35 6.9E+04EDIP 2003 combined [kg NO−3 ] 3.6E-02 1.3E-02 2.2E+01 0.35 7.5E+03TRACI [kg N] 2.7E-04 9.5E-05 1.2E-01 0.35 1.0E+06Ozone depletion potentialCML 2001 [kg CFC-11-eq] 7.5E-08 2.1E-08 2.7E-05 0.28 3.6E+09EDIP 2003 [kg CFC-11-eq] 7.5E-08 2.1E-08 2.7E-05 0.28 3.6E+09TRACI [kg CFC-11-eq] 7.1E-08 6.4E-09 2.4E-05 0.09 3.8E+09Table 2.2: Selected LCIA results for selected processes modeled in ecoinvent V2 databasethe relative difference between one desktop computer and one kg of steel can vary by several orders ofmagnitude across different LCIA schemes. This result suggests that impact scores calculated in one ofthese dimensions have internal validity only; one cannot compare acidification scores (e.g.) from twodifferent LCIA methodologies even in a relative sense. This is not necessarily problematic, so long asnumerical impact scores are used within a consistent framework, but it imposes an additional constraintson the appropriate use of these scores outside of their original source study contexts, and makes suchimpact dimensions unsuitable for integrative studies which must incorporate LCA results from multiplesources, except in circumstances where the modeling assumptions across sources are guaranteed to beequivalent.Table 2.3 shows the results of LCIA calculations for select substances for a similar set of impactcategories using both the ecoinvent V2 and GaBi 4 databases. The table shows the ratio of the GaBioutput to the ecoinvent output; a result of 1.0 indicates the two outputs were identical. GaBi provideslarge-scale empirically-driven industrial models that are very similar to those of ecoinvent, but it is a31competitor product and thus distinct. The substances modeled are nominally equivalent, according totheir descriptions, but the internal modeling assumptions and data sources may vary considerably. Thiscomparison shows the degree to which these two databases are internally consistent.Electricity Steel Passenger Circuitcar boardGlobal warming potentialCML 2001 100-yr [kg CO2-eq] 0.99 1.01 1.04 1.22EDIP 2003 100-yr [kg CO2-eq] 0.99 1.03 1.04 1.22TRACI [kg CO2-eq] 0.98 1.02 1.04 1.22Acidification potentialCML 2001 global [kg SO2-eq] 1.09 0.74 0.23 1.76EDIP 2003 [m2] 1.05 0.69 0.18 0.12TRACI [moles H+-eq] 1.08 0.75 0.23 1.76Eutrophication potentialCML 2001 global [kg PO4-eq] 0.03 0.11 0.26 0.14EDIP 2003 combined [kg NO−3 ] 0.96 0.82 0.63 1.45TRACI [kg N] 0.01 0.60 0.09 0.71Ozone depletion potentialCML 2001 [kg CFC-11-eq] 4.30 0.47 N/A 0.39EDIP 2003 [kg CFC-11-eq] 3.85 0.47 N/A 0.39TRACI [kg CFC-11-eq] 4.35 0.56 N/A 0.47Primary energy demand [MJ] 0.98 0.94 1.05 0.77Electricity, GaBi: Power grid mix, AC consumption mix, at consumer 230V, Germany [kWh]Electricity, ecoinvent: electricity, low voltage, at grid, Germany [kWh]Steel, GaBi: DE: Steel sheet (ECCS) BUWAL [kg]Steel, ecoinvent: sum of: Steel, low-alloyed, at plant; Sheet rolling [kg]Passenger car, GaBi: technology mix, gasoline driven, Euro 4, passenger car [vkm]Passenger car, ecoinvent: operation, passenger car, petrol, EURO4 [vkm]Circuit board, GaBi: avg. of 4-layer and 8-layer printed wiring board rigid FR4 with HASL finish [m2]Circuit board, ecoinvent: avg. of lead-free/leaded printed wiring board, surface mount (6-layer HASL) [m2]Table 2.3: Ratio of modeled outputs from GaBi 4 relative to ecoinvent v2, selected processes andimpact categoriesThe first three substances, electricity, steel production, and passenger car transport, are relativelysimple and dominated by energy flows, while the fourth, a circuit board, incorporates many componentsand high-complexity materials; thus modeling uncertainty could be expected to be smaller for the firstthree substances, which is borne out in most cases. The results for global warming potential are fairlystable, within 5% for the first three substances, and a 22% gap for the circuit board; primary energydemand is similarly stable. Other categories vary widely. The LCIA characterization schemes are sup-32posedly implemented in identical ways, and the underlying models should be very similar, accordingto the database documentation for each. Nevertheless, in some categories the results differ by an orderof magnitude or more. This suggests that the LCIA calculation has a very high sensitivity to pertur-bations in the underlying input models; the differences between the models is nominally small, but thedifferences between the results are large. This result may be compared to a study of the uncertainty incalculating device carbon footprints, which used a server as a case study, and compared impacts usingprocess-sum methods based on both ecoinvent and GaBi databases, as above but including a much moredetailed adjustment to produce equivalent models [133]. The results of that analysis suggested a 95%confidence interval of ±15% relative to the mean for estimates of global warming potential of the pro-duction phase of ICT devices; this range includes model uncertainty relating to data inputs, but excludestruncation error, which is discussed below.The stability of global warming potential and primary energy demand can be explained by the rela-tive simplicity of these impact categories, as they involve the aggregation of a relatively small numberof flows, and possibly by the relatively large amount of physical science and empirical research resultsrelating to these categories. On the basis of these results, the research in this dissertation considers onlyglobal warming potential, and in the case of Chapters 3 and 4, primary energy demand. Notwithstand-ing the relative stability of these categories, care must be taken when integrating results from differentstudies which may rely on different underlying data sources, especially for complex substances likeelectronic components.There are many significant technical issues in LCA regarding the ongoing development of charac-terization factors, definition of impact categories and aggregation schemes, and linkage of impact scoresto downstream damage categories. A summary of technical issues is provided by Reap et al [134, 135].Given the empirical evidence above that global warming potential and primary energy are relatively sta-ble across different LCIA schemes and modeling databases, this dissertation does not further considerthese technical issues.2.4.2 Truncation errorProcess-sum LCA is vulnerable to cut-off or truncation error due to components of the system notincluded within the study boundary. It is not possible for a process-sum model to precisely include allimpacts while maintaining a finite study boundary; thus a fundamental assumption of LCA is that it is33possible to define study boundaries such that the study reasonably represents the impacts of the system.A key assumption of most process-sum studies is thus that impacts from outside the study boundarywould, if accounted for, not fundamentally change the conclusions of the study. This is more likely tobe true if the study attempts to identify an environmentally preferable alternative within a comparativeframework, rather than an absolute numerical impact score.In some cases, it is possible to obtain the impact of a system from the top down, by starting from alarger known impact and determining the share of that impact which should be allocated to the systemin question. Economic input-output LCA (EIOLCA) has emerged as a prominent top-down method forLCA which determines allocation according to economic flows (i.e. in monetary units) using economicinput-output tables [80, 136]. EIOLCA is a coarse methodology, well-suited to the study of industriesand economic sectors but less so to specific devices, but sidesteps the issue of truncation error [79].Hybrid methods have been developed which begin with a process-sum LCA result and adjust it upwardsusing economic data to correct for truncation error [137]; these are considered by many experts to bethe most accurate means of estimating device footprint [81, 138, 139]. However, they do impose higherdata requirements than process-sum LCAs.Prior work has demonstrated the importance of truncation errors. In particular, when comparingsystems which have different value chains involving different industrial mixes, the ranking of thosesystems in terms of impact may be determined entirely by the study boundaries and corresponding trun-cation error, rather than by the fundamental physical processes associated with those systems [81, 140].Past hybrid LCA studies have quantified truncation error as ranging from 20% up to 80%, with moststudies reporting results between 30% and 50% [81]. Of the two hybrid LCA studies of ICT equipment,the first, a study of a desktop, estimated that process-sum LCA accounted for 50% to 60% of the totalimpacts of the production phase [13], while the second, a study of a laptop, estimated that process-sumLCA accounted for 44% of total impacts of the production phase [92]; the remainder of impacts werecalculated using an economic correction based on EIOLCA methods. Given the potentially large mag-nitude of truncation error, LCA studies should use hybrid analysis where possible, and especially whendetermining numerical impact results which are likely to be applied in contexts external to the study,such as device carbon footprints.A major barrier to the use of hybrid LCA is the added complexity it brings. The required com-binations of process-sum and EIOLCA databases is intricate, and brings added data requirements of34component pricing and industrial economic data, to which the output results are sensitive [138]. In therecent hybrid study of a laptop PC, the economic correction was derived using estimates of the value ofcomponents and raw materials in the laptop, along with its purchaser price [92]. While this method maycorrect for truncation error, it introduces new sources of error due to uncertainty in component and prod-uct prices, which are typically trade secrets, especially for newer products. Prices can be particularlyvolatile for rapidly emerging and evolving products.The study in Ch. 4 considers impact of ICT devices using a process-sum approach, rather than ahybrid approach. Difficulty in obtaining reliable pricing data along with uncertainties related to theintricacy of hybrid LCA was a key consideration in this methodological decision. Of equal importance,however, was the desire to be able to benchmark and compare the results against those of the ecoinventdatabase, which were derived using process-sum LCA, as well as enabling comparisons with otherpublished LCA studies of ICT devices, the vast majority of which were obtained using process-sumLCA, including Apple’s extensive product environmental report dataset. The presence of truncationerror in Ch. 4 means its results can be used for internal comparisons between devices only, rather thanas device footprints which might be applied outside of the study’s modeling framework.The study in Ch. 5 is not a traditional LCA in that it does not specify device and system inventories,but rather combines existing results using an LCA approach. It is still vulnerable to truncation errorwhich may exist in the integrated studies. However, that error should not greatly affect the conclusions,for two reasons. First, a significant portion of the integrated studies were themselves performed usingEIOLCA or similar top-down approaches, and thus should have accounted for truncation error already.Second, the majority of impacts in the systems under study are due to operational electricity use, whichare characterized by a grid emissions factor that has been extensively studied and is widely used in manyresearch domains.2.4.3 Toxicity and environmental healthThe framework applied throughout the research in this dissertation is well-suited to a conception ofimpacts that is quantitative and expressible in units amenable to linear modeling such as greenhousegas emissions in kg CO2e. However, some important types of impacts of ICT are less well-suited toquantitative linear modeling, or to quantitative expression in any terms. While these impacts are beyondthe scope of the research in this dissertation, they are briefly discussed here for completeness.35ICT devices can contain toxic substances which may be harmful to humans and ecosystems. Lifecycle assessment can be used to assess relative human health and ecosystem damage potentials, butLCA models are designed to estimate average potential impacts of typical processes at large scales,e.g. continental or global scales in the case of the USEtox model which underpins many LCA impactaggregation schemes [141]. Informal e-waste recycling operations in lower-income countries have beenshown to result in release of toxic substances, including heavy metals, leading to environmental contam-ination and adverse human health outcomes [142, 143, 144, 145, 146, 147, 148, 149, 150]; these acute,localized impacts are primarily assessed through the use of field measurements and are not incorporatedinto current LCA models.Impacts due to toxic substances in electronics have been conceptualized using a red-list approach,exemplified by the EU’s Restriction of Hazardous Substances (RoHS) Directive [151] which restrictsthe use of lead, mercury, cadmium, hexavalent chromium, polybrominated biphenlys, and polybromi-nated diphenyl ether; similarly, the EPEAT voluntary labeling scheme that assigns points for eliminationof the same list of substances [35]. In addition, e-waste management schemes, exemplified by the EUWaste Electrical and Electronic Equipment (WEEE) Directive [152] which sets targets for collectionand recycling of waste electronics in the EU, are a key mitigation strategy for reducing opportunitiesof toxic releases. There remains an ongoing need for primary research to identify high-priority toxicsubstances, alongside regulatory structures to encourage the elimination of such substances from elec-tronics devices and to reduce the risk of release of toxic substances to environment through improvedwaste management.2.4.4 General bounds on LCA resultsLCA studies rely on models of technological and natural processes to approximate the potential envi-ronmental impacts of a device or system. There are a few caveats when considering the validity andinterpretation of these results. First, unless specifically otherwise modeled, models of upstream pro-cesses are intended to be representative of typical or average industrial conditions. Obtaining actualemissions flows from specific physical facilities is beyond the scope of most LCAs, especially if LCAdatabases are used. Thus, the impact scores should be thought of as average representational impacts,rather than specific physical impacts. For this reason, the impact dimensions represent potential impacts(e.g. global warming potential), rather than physical emissions. Scientific consensus defines these po-36tential scores to be meaningful and reasonable expressions of likely impacts, but they are synthetic, notempirical, and only as accurate and meaningful as the underlying models and data.In addition, the methods discussed thus far have followed the attributional style of LCA, which isuseful for defining the impacts arising from industrial processes related to a device or system. Broaderimpacts on society, such as changes in consumption patterns, are not captured in attributional LCA.There are other methods available, such as consequential LCA, which can measure the effects of largescale industrial changes, e.g. due to a policy; these involve economic frameworks and have very differentdata requirements and methods [79]. Consequential LCA is not used in this dissertation or in anyexisting LCA studies of the impacts of ICT devices and services. One should not attempt to infer largescale societal changes from attributional LCA alone.Finally, notwithstanding the apparent numerical instability of impact models for dimensions otherthan global warming potential and primary energy demand, restricting the definition of impact to thosetwo dimensions is potentially problematic, as they may not capture many important impacts. For exam-ple, the global warming potential of the end-of-life phase of electronics is very small [1], but e-wastelandfill and recycling operations may be very damaging to environment and human health, especially ina lower-income informal context. A lower GWP score does not necessarily mean a lower environmentalimpact; it only means a lower impact in terms of global warming potential.2.4.5 Alternative approachesLCA is one of a family of quantitative approaches towards environmental impact assessment. Othermethodologies exist, such as ecological footprint analysis, which aggregates impacts in terms of equiva-lent land area usage [153]; ecosystem services assessment, which aggregates impacts according to theireffect on valuable properties of ecosystems, e.g. their ability to provide clean air and water; and materi-al/substance flow analysis [154] and structural path analysis [155], which quantify stocks and flows ofmaterials through economies and ecosystems. Each methodology comes with its own set of limitationsand challenges; none of them have seen any appreciable use in the context of studying environmentalimpacts of ICT devices and services.One possible approach towards simplified analysis of ICT devices and services is to consider op-erational energy use alone. There is considerable value in understanding the way energy is used, forexample in order to aid the design of policies to encourage demand-side reductions in energy use, and37there is a rich history of literature regarding the energy usage of ICTs (see Section 2.2.1). Some studiesof ICT services have chosen to study operational energy only [41], or global warming potential dueto operational energy only [45], excluding embodied impacts due to devices and infrastructure. Thisgreatly simplifies the analysis, but of course significantly narrows the scope of the study as well; theappropriateness of this decision depends on the research goals of the study.2.5 Chapter summaryRegarding environmental impacts of ICT, knowledge gaps have arisen due to rapid change in the indus-try, such as the introduction of new devices enabled by Moore’s Law, convergence in broadcast TV /internet services, roll-out of high bandwidth mobile networks, and growth of cloud computing services.LCA and operational energy accounting were identified as useful methodologies for estimating suchimpacts using quantitative linear models; both have seen substantial use in existing literature. Specificknowledge gaps addressed in this dissertation are ambiguity in the existing literature of LCAs of desk-top PCs (Chapter 3), lack of coverage regarding impacts of newer devices along with high time anddata costs of conducting LCAs for such devices (Chapter 4), and lack of high-level forward-lookingintegrated views of modern network-enabled services (Chapter 5). Impact dimensions were limited toGHG emissions and primary energy demand due to apparent numerical instability of models of otherimpact dimensions.38Chapter 3Meta-analysis of LCAs of desktop PCs3.1 IntroductionInformation and Communication Technology (ICT) equipment has come under scrutiny in the pastdecade due to rising concern over their environmental impacts. According to the best recent estimates,the global ICT sector is responsible for about 2% of global anthropogenic greenhouse gas emissions;half of these emissions are due to Personal Computers (PCs) and peripherals [30]. Quantifying theenvironmental impacts of PCs is important in order to understand both the aggregate impacts of theindustry, and the relative share of the impacts from residences and offices that can be attributed to PCs.Several life cycle assessment studies have focused on consumer electronics including PCs. Unfor-tunately, the results of these studies are not consistent with one-another, with significant variation inboth the absolute impact reported, and the life cycle phase that dominates the impact, which is eitherthe manufacturing or use phase. This literature has been recently reviewed by several authors includingJames and Hopkinson [156], Malmodin et al [25], Yao et al. [20], and, most thoroughly, Andrae andAndersen [19]. Each review noted the lack of consensus in the literature, but none attempted to providea systematic and rigorous exploration of the source of the disagreement, or to identify which (if any) ofthe reviewed studies might be the most accurate, a task which is undertaken here. Focusing specificallyon desktop PCs without displays in order to limit variation, existing LCA studies are disaggregated intoimpact assessments for each life cycle phase; data sources used to assess for each phase were identi-fied. The validity of all reported data sources and modeling assumptions is assessed. This work hasa significantly narrower scope than Andrae and Andersen and so allows for greater depth in order to39improve understanding the results of existing studies of desktop PC impacts. Associated literature isalso surveyed which helps contextualize and interpret these results.First, relevant studies are summarized in order to show their reported results in terms of total en-vironmental impact, and relative impact by life cycle phase, and to identify which components of thedesktop PC’s life cycle inventory are the largest contributors to the total impact of each life cycle phase.The study then addresses the following two questions:1. To what extent and for what reasons do estimates from different studies disagree with one another?2. Given the uncertainties in estimates for the environmental impacts of a desktop PC, and in lightof the various studies, what are the best guess estimates or ranges for evaluating the impacts ofdesktop PCs?The study’s approach to structuring the analysis and answering these questions is described in thefollowing section.3.2 ApproachMeta-analysis of LCA studies is challenging, because studies can differ in terms of unit of analysis,temporal and geographic scope, inventory data, impact factor data, and impact characterization scheme.These differences each add variation to the numerical result and must be accounted for where possiblein order to facilitate a valid comparison.3.2.1 Scope and unit of analysisThe study focuses on desktop PCs, excluding other electronic devices like laptops and cell phones,because desktops have received the most attention to date from LCA researchers, and because theyrepresent a relatively large share of the total impact of the ICT sector. Desktop PCs typically include aCRT or LCD display, peripherals such as a keyboard, mouse, and printer, and a central unit, often calledthe control unit. The scope of comparison in this study is restricted to control units only, excludingdisplays and peripherals. Displays in particular are known to represent a significant component of thetotal impact of a desktop PC system, but they are not treated consistently in the literature, with somestudies assuming CRT displays, some assuming LCDs, many assuming a mixture of the two, and someexcluding displays altogether. An accurate comparison of these studies would have to separate the40display and control unit and compare them independently, and while a comparison of LCA estimatesof displays would be valuable, such an analysis is not performed here in order to limit the scope andcomplexity of this study. As such, when the study refers to desktop PCs, this should be understood torefer to the control unit only. Desktop PCs come in different sizes and with different features, but moststudies aim to measure a PC with characteristics representative of an average or typical product withinthis category; this study follows the same approach.Geographic scope does vary from study to study; manufacturing impacts usually occur in southeastAsia, but use-phase impacts depend on the local electricity generation mix. To correct for this sourceof variation, use-phase impacts are compared in terms of kWh of electricity used rather than endpointimpacts. Temporal scope also varies; results are reported according to their year of publication with aslight preference given to more recent studies when weighing evidence, as they are more likely to berepresentative of modern products.3.2.2 Definition of environmental impactImpact can be measured in many different ways depending on the life cycle impact assessment (LCIA)method chosen by the practitioner. Two categories, global warming potential (GWP100, in kg CO2e,per the Intergovernmental Panel on Climate Change (IPCC) standard) and cumulative primary energydemand (CED, in MJ), were used in a large number of studies. Some categories like ecotoxicity andeutrophication potential were used only in a few studies, and in some cases were measured with differentunits, making a comparison across studies in these categories much more difficult. This study thereforereviews reported impact measurements in both global warming potential and primary energy demand.3.2.3 Analytical approachThe central challenge is judging the quality of the LCA studies and commenting on their reasonable-ness. This study’s strategy for doing so is decomposition: reported results are split first into componentlife cycle phases; second, into the largest sources of impacts for each phase (that is, decomposing themanufacturing phase into a listing of key modules such as mainboard, integrated circuits, and so on);and third, into an inventory listing of items, in kg for material parts and kWh for electricity, and a corre-sponding impact factor, in MJ per item unit and kg CO2e per item unit. Assumptions in each study arecompared at this fine-grained level, highlighting key differences that are obscured in the total result.41In judging reasonableness, a non-quantitative weight-of-evidence approach is employed. It is oftenvery difficult to independently verify reported results of life cycle assessments, but some aspects ofthe desktop PC life cycle, notably use-phase electricity usage and product lifespan, have been studiedoutside the body of LCA studies surveyed here. This study surveys this literature as well, assigns a dataquality score to each study, and assumes that reasonable representative values lie towards the middle ofthe range reported by the highest-quality evidence; comparing against this broader context allows for arough judgment as to whether the LCA studies make reasonable assumptions. For parts of the desktopPC life cycle without a large ancillary literature, it is more difficult to judge reasonableness; instead thestudy identifies clear outliers or errors, and suggests that more weight should be given to higher-qualitystudies, as demonstrated by thorough reporting of methods, data sources, and results. Recognizing thelimitations of this approach, the study draws only cautious conclusions in such cases. In all cases, dataheterogeneity precludes a rigorous statistical aggregation with probability distributions. Instead, thestudy assigns reasonable upper and lower bounds on impact measurements.In answering these questions, the study attempts to mitigate some of the uncertainty surroundingdesktop PC LCAs, and provide the reader with appropriate information and context needed to evaluatesuch efforts, and to accurately interpret their results.3.3 List of studies and their overall resultsLCA studies of desktop PCs were identified through literature searches of Google Scholar and a numberof academic databases, and through citations in other published studies and reviews. A 1993 studyis excluded due to its age [157], and one study [158] because the same results were re-published inanother [159]; to our knowledge, all other published LCAs of desktop PCs are included in this review.Details about the scopes and methodologies of these studies used for this work are as follows, with thestudies identified by their authors:• Tekawa et al, published in 1997 [160], is a process-sum study, based in Japan, which focuses ondesktop and laptop PCs and reports several types of impacts, but includes minimal informationabout the unit of analysis and does not provide a bill of materials. There is not enough informationto separate the CRT display impacts from the control unit impact, so the total impacts in this studyinclude the display.42• Atlantic Consulting, published in 1998 [161], was a major process-sum study for the EuropeanUnion’s EcoLabel scheme and provides a full listing of impacts, though the bill of materials is notfully specified, making it impossible to split the impact into components.• Williams, published in 2004 [13], uses a hybrid economic input-output LCA methodology, withthe intention of correcting for truncation errors inherent in process-sum methods relating to man-ufacturing impacts.• Masanet et al., published in 2005 [162], from Lawrence Berkeley National Laboratories and pre-pared for the California Energy Commission, calculates impacts for desktops manufactured andused in California using an EIOLCA approach. Manufacturing data is adapted from earlier studiesfrom Williams [13, 163], and energy data is taken from Kawamoto et al. [164].• Hikwama, published in 2005 [165], is an undergraduate thesis from the University of South-ern Queensland. Characterization factors are taken from a proprietary database, but the studycontributes an inventory which was created through disassembly and manual weighings. Theinventory was adapted for use by Eugster et al. [158] and Duan et al. [159].• Kemna, published in 2005 [166], known as the Methodology Study for Eco-design of Energy-using Products, or MEEUP, was conducted for the European Commission by VHK and studied anumber of energy-using electrical products, including computers, using a process-sum approach.The study contains a detailed inventory and a particularly strong meta-analysis assessing use-phase energy consumption.• Choi et al., published in 2006 [167], based in Korea, examined several different impact cate-gories and focused in-depth on the influence of recycling on total impact, but provides minimalinformation regarding the inventory or impact factors assumed in the analysis.• Braune and Held, published in 2006 [168], known as the Development of Environmental Perfor-mance Indicators or EPIC-ICT project, conducted for the European Commission and co-ordinatedby researchers at IKP Universita¨t Stuttgart, contains detailed case study process-sum LCAs ofdesktop PCs, but does not provide the full inventories or the absolute impact results; instead theproject focused on identifying areas of leverage for future eco-design initiatives.43• Hischier et al., published in 2007 [95], describes the studies in the ecoinvent database, which isavailable in the unit process titled “Desktop computer, without screen, at plant/GLU U”. This isthe most transparent of the surveyed studies. The desktop PC data comes from a mixture of pri-mary measurements and imported results from Kemna [166] and Braune and Held [168]. Primaryenergy and GWP results were obtained from the ecoinvent database using the IMPACT2002+methodology.• IVF, published in 2007 [169], was conducted for the European Commission’s Energy UsingProducts (EuP) program, and focused on identifying eco-design requirements. Like the MEEUPproject, this report contains a strong review of previous literature. The LCA study within is quitetransparent and contains a full bill of materials and listing of impacts.• Duan et al., published in 2009 [159] is apparently a re-publication of results also available inEugster at al [158]); both studies examine the life cycle impacts of a PC manufactured and used inChina, partially adapting data from the ecoinvent database and from Hikwama [165]. Results areprovided as aggregated Eco-Indicator points which unfortunately makes it difficult to determinehow the results were obtained or how they might compare with other studies.• Apple product environmental reports, published in 2010 [170, 171], are environmental declara-tions for the Mac Mini and Mac Pro, respectively, publicly available on Apple’s corporate website.These studies provide estimates of the global warming potential of each product disaggregated bylife cycle phase, but provide no methodological details as to how the figures were reached.Results of studies included in this synthesis are shown in Figure 3.1, where the left-most axes showthe relative impacts of each phase of the life cycle, and the right-most axes show absolute impacts, whereavailable. Numerical data tables for this and other figures are in Appendix A. Studies that measuredimpact in terms of cumulative primary energy demand in MJ are on the top; studies that measuredglobal warming potential in kg CO2e are on the bottom. Where possible, the results have been adjustedto exclude displays. Only one study, Tekawa et al [160], did not provide enough information to do so;its results thus include a CRT monitor. All other studies report the desktop PC control unit only.The relative impacts on the left side of the figure suggest that the use phase is the dominant compo-nent in total impact, according to most studies. The studies measuring primary energy on average report440 50 100Relative impact by phase (%)−10Atlantic Consulting 1998Williams 2004Kenma 2005 [MEEUP]Masanet et al 2005Braune and Held 2006 [EPIC−ICT]Hischier et al 2007 [ecoinvent]IVF 20070 50 100−10Atlantic Consulting 1998Tekawa 1998 (includes CRT)Hikwama 2005Masanet et al 2005Braune and Held 2006 [EPIC−ICT]Choi 2006Hischier et al 2007 [ecoinvent]IVF 2007Duan 2009Apple 2010a (Mac Pro)Apple 2010a (Mac Mini)0 5000 10000 15000 20000Total absolute impact0 500 1000 1500ManufacturingDistributionUseEnd of lifePrimaryenergy(MJ)Globalwarmingpotential(kg CO2e)Figure 3.1: Summary of LCA studies showing breakdown of life cycle energy use and carbonequivalent impactsuse phase impacts of 62% of the total, with 35% due to manufacturing and very small portions for distri-bution and end-of-life. Studies measuring global warming potential similarly report an average of 58%for use phase, 39% for manufacturing phase, and very small portions for distribution and end-of-life.End-of-life and distribution are not considered in this analysis because of their relatively small im-pacts in terms of MJ and kg CO2e. Small impacts in terms of primary energy and global warmingpotential should not however be taken as indications that the end-of-life life-cycle phase is environmen-tally benign, as impact categories relevant to toxic releases may show higher scores for the end-of-lifephase. In particular, informal recycling in lower-income countries is not captured in LCA databases andis difficult to accurately characterize, but has been shown to have significant negative health impacts onworkers and local residents due to environmental contamination; see [21] for a review on this topic.The remainder of this study focuses on the manufacturing and use phases, and attempts to deducethe reasons underlying the disagreements across the studies shown in Figure 3.1. In particular, the studyfocuses on understanding why studies that report dominant impacts from the manufacturing phase,namely those of Williams [13], Masanet et al [162], and Choi et al [167], disagree with the majority ofthe LCA studies that show the use-phase as dominant.453.4 Manufacturing and production phasesThe manufacturing phase of the life cycle, which includes material extraction and processing, sub-assembly production, and final assembly, is particularly difficult to analyze because of the large numberof highly complex processes involved. The analysis can nonetheless be made tractable by examiningthe impacts of a relatively small number of component categories. Figure 3.2 shows the results of aninitial review, in which the impacts have been divided into component categories: mainboard-integratedcircuits (ICs), mainboard-other, power supply, hard drives and disk drives, and the metal or plastic cas-ing. Graphics cards and other internal circuit-board-based peripherals are grouped with the mainboard.All additional items are grouped as “other”. For details on how the source study’s results were adjustedin order to fit into these component categories, see Appendix A.Figure 3.2: Manufacturing impacts of desktop control unit components (relative)The study goal, to identify the quantity and impact per quantity unit of each component category, ischallenged by a lack of published complete bills of materials or complete listings of impact factors, andby differences in the way the respective authors structure their analyses. For example, some studies con-sider ICs separately and others lump them in with the mainboard. Nevertheless, as shown in Figure 3.2,it is possible to make observations about relative impacts of component categories: the mainboard in-cluding ICs is responsible for the largest impact, accounting for more than 50% of the impacts in all butone study, with the other components constituting small but non-negligible proportions of the remain-der. The proportions vary considerably, as do the absolute totals of the production phase visible on theright half of Figure 3.1; in order to explore the sources of this variation, the quantity and impact of eachcomponent category is identified in the following sections.463.4.1 Component-level impactsThe study’s approach here is to identify the quantity, in kg, of each component category, and its impact,in terms of primary energy, in MJ per kg of component, and global warming potential, in kg CO2e perkg of component. Figure 3.3 shows results of our survey. Integrated circuits are excluded from thisfigure because they have very high impacts and negligibly small mass, making a comparison inappro-priate; they are examined separately in the following section. Product packaging such as cardboard orpolystyrene foam is excluded from the total product mass. Note that only four full inventories for com-ponent categories were available – Williams [13], IVF [169], Hischier [95], and Hikawama [165]. Thestudy by Williams does not provide global warming potential; the study by Hikwama provides neitherglobal warming potential nor primary energy.024681012Total desktopCasingMainboardDrivesPower supplyPart massper desktop (kg)01020304050607080Total desktopCasingMainboardDrivesPower supplyGHG per part(kg CO2e/kg)0200400600800100012001400Total desktopCasingMainboardDrivesPower supplyPrimary energyper part (MJ/kg)  Williams 2004IVF 2007Hischier et al 2007Hikwama 2005Figure 3.3: Mass and impacts of desktop PC components (excluding mainboard ICs)The studies roughly agree over the mass of the components, with no obvious outliers; yet theydisagree on the impacts per part. Hischier’s results are consistently much larger than IVF’s, roughly bya factor of two. Thus, for apparently similar inventories, the two studies report significantly differentfinal results. The primary energy impacts from Williams vary significantly from both of these studies.With a limited number of data points no specific conclusions can be drawn except that the impacts maylie in the range shown on these graphs. Ideally it would be possible to benchmark these results againstother external studies, as is done later in this study for use-phase energy consumption, but we are notaware of any such studies focusing on these particular component categories. Fortunately, the literature47for integrated circuits is slightly richer, which is drawn upon in the next section.3.4.2 Integrated circuitsThe LCA studies measure semiconductor content in three ways: in area of input silicon wafer; area offinished silicon die; and mass of packaged chip. In the latter case, a mass ratio of finished silicon dierelative to the mass of the packaged chip is assumed. It is possible to convert between these units bymaking a few assumptions; calculations described Appendix A result in a silicon wafer input of roughly1.67 cm2/cm2 finished die, and a mass of 0.2 g/cm2 finished die. Table 3.1 shows the semiconductorcontent per desktop according to each of the studies that reported it. The area of the table labeled“adjusted inventory” shows this study’s estimates of the total mass and area of finished die given thereported inventories and these calculated conversion factors.Williams(2004)Yaoetal.(2010)Hischieretal.(2007)Kemna(2005)IVF(2007)Williamsetal.(2002)Krishnanetal.(2008)Boydetal.(2009)AndraeandAnderson(2010)Reported inventoryInput wafer (cm2/desktop) 110 – – – – – – – –Finished die (cm2/desktop) – 12 – – – – – – –Packaged chip, 0.9% Si (g/desktop) – – 30 – – – – – –Packaged chip, 1% Si (g/desktop) – – – – 95.5 – – – –Packaged chip, 2% Si (g/desktop) – – 59 – – – – – –Packaged chip, 5% Si (g/desktop) – – – 100.4 69 – – – –IC primary energy (MJ/desktop) 1992 99 1175 – 463.6 – – – –IC global warming (kg/desktop) – – 72.4 – 34.8 – – – –Adjusted inventorySi mass (g/desktop) 13.2 2.4 3.9 5 4.4 – – – –Finished die (cm2/desktop) 66 12 19.4 25.1 22 – – – –Impacts per finished die areaPrimary energy (MJ/cm2) 30.2 8.3 60.6 – 21 31 33 81 –Global warming (kg CO2e/cm2) – – 3.7 – 1.6 – – 5.5 7Table 3.1: IC inventories and impacts per desktop mainboard, with originally reported inventoryand adjusted inventory assuming 0.2g/cm2 finished dieThe table also shows the impacts per finished die area, measured in MJ/cm2 and kg CO2e/cm2. For48the five LCA studies, the figure is obtained by dividing the total reported IC impact, where available,by the adjusted finished die area. In addition, three studies were identified that assessed the life cycleimpacts of semiconductors alone; these results are included on the right side of the table. The latter,from Boyd [172], reports a total of 91 MJ per die and 6 kg CO2e per die for all production phases,which corresponds to 81 MJ/cm2 and 5.5 kg CO2e/cm2 assuming an average 1.11 cm2 die in a 45 nmprocess, according to their data. Likewise, an algorithm from Andrae and Anderson [93], based inpart on Boyd’s work, estimates 34.7 kg CO2e/g of die, which would be 7.0 kg CO2e/cm2 assuming thisstudy’s estimated conversion rate of 0.2 g/cm2.Overall, the finished die area ranges from 12 cm2 to 66 cm2 per desktop PC - though the latter figurewas criticized by Yao as being an overestimate [20] - with a median of 22cm2. There is evidence thatYao’s impact estimate is too low due to an apparent misapplication of Williams’ methods, and Hischier’sis too low due to an apparent error in calculating the die area of a packaged chip; see Appendix A fordetails. Discarding these, reported primary impact estimates range from is 21 to 81 MJ/cm2, and globalwarming impact estimates range from 1.6 to 7.0 kg CO2e/cm2. Boyd’s study [172], which reportedimpacts at the upper end of these ranges, is the most thorough and up-to-date LCA of semiconductormanufacturing, which suggests that other studies may be underestimating impacts due to semiconduc-tors. Establishment of standard impact factors and inventory reporting schemes would help a great dealin removing some of this uncertainty.3.4.3 Total manufacturing impact: analysisStudies at the level of component categories disagree significantly. For the component categories shownin Figure 3.3, inventories are relatively consistent, indicating that much of the variation is due to differentassumptions regarding the impacts of the various components, especially the mainboard and ICs. Forintegrated circuits, both inventory and impacts were highly variable. Using the data from Figure 3.3,there is a range of total desktop mass from 9.0 to 11 kg; a global warming potential ranging from 13 to23 kg CO2e/kg desktop, and a primary energy consumption ranging from 180 to 590 MJ/kg desktop.The data is not sufficient to determine a plausible range for the manufacturing impacts of a desktopPC. The data can be summarized by assuming that both total desktop mass and impacts per desktopmass may fall within the ranges reported in Figure 3.3; thus by multiplying these ranges together, thetotal global warming potential could vary from an estimated 120 to 250 kg CO2e/desktop, and the total49primary energy consumption could vary from 1600 to 6500 MJ.3.5 Use phaseThe use phase is responsible for the highest impact according to a majority of studies as shown inFigure 3.1. Impacts due to the use phase are simpler to measure and quantify than impacts in the man-ufacturing phase, as the only impacts are due to electricity consumed by the device during its lifespan.Nevertheless, the LCA studies vary to a surprising degree, with estimates of lifespan primary energyconsumption due to the use phase ranging from a low of 580 MJ (Williams [13]), to a high of 16800 MJ(Kemna [166]), which is a 30-fold variation.Total use-phase electricity consumption is a function of the power demand of the device, patternsof usage, and the lifespan of the device. Most devices have several different operating modes, suchas active, standby, or off, which have different power demands. Using methods described by Roth etal. [82] and Kawamoto et al. [164], given a set of power modes PM such that, for mode i ∈ PM, if theaverage power draw in mode i is Pi, and the average time spent in mode i is ti, then the total lifespanenergy consumption can be found using the following equation:Lifetime energy[kWh] = Lifespan[years] · ∑i∈PMPi[kW] · ti[hoursyear](3.1)The total of the summation on the right side of the equation is sometimes called the Unit EnergyConsumption (UEC), measured in kWh/year. Variation between studies occurs due to differing estimatesfor the lifespan, power draws, and time share which cause variation in the UEC. In addition, whenenergy consumption is converted to an impact such as global warming potential or primary energydemand, assumptions regarding the characterization factors may vary across studies as well. Each ofthese causes of variation is analyzed in turn below.3.5.1 Unit energy consumptionIn addition to the LCA studies already introduced, a significant body of literature has assessed the energyconsumption of consumer electronics. Results from these studies are summarized in Figure 3.4, withthe annual unit energy consumption in kWh/year on the y-axis (desktop control unit only, no display),and the year of publication on the x-axis. Studies have been assigned a data-quality score, with large-nprimary studies and meta-analyses marked “high”, smaller or less rigorous studies marked “medium”,50and n = 1 studies or assumptions marked “low”; the size of the marker corresponds to the data quality.The marker shape indicates whether the study measured use in a home setting (square), an office setting(circle), or both/not specified (triangle). Primary studies are shaded dark.2000 2005 2010010020030040050060070080012345 6 78 91011 1213141516171819 2021 2223242526 272829303132Unit energy consumption (kWh/year)Year of publicationDesktop PC use−phase energy (no display)Usage patternHome useOce useBothStudy typePrimarySecondaryMixtureData QualityLowMediumHigh1 [82] Roth et al. (2002) (high-power PC): Secondary data2 [168] Braune (2006) [EPIC-ICT] (gamer PC): Dell data3 [173] US EPA (2010) [Energy Star, baseline]: Secondary data4 [174] US EPA (2005) [Energy Star, baseline]: Secondary data5 [168] Braune (2006) [EPIC-ICT]: US EPA data6 [166] Kemna (2005) [MEEUP]: Meta-analysis of prior studies7 [175] Moorefield (2008): California plug-load audits8 [82] Roth (2002): Secondary data9 [174] US EPA (2005) [Energy Star, baseline]: Secondary data10 [166] Kemna (2005) [MEEUP]: Meta-analysis of prior studies11 [168] Braune (2006) [EPIC-ICT]: Dell data12 [173] US EPA (2010) [Energy Star, certified]: Secondary data13 [174] US EPA (2005) [Energy Star, certified]: Secondary data14 [176] Porter (2006): California plug-load measurements15 [95] Hischier (2007) [ecoinvent]: Secondary data16 [4] Roth (2007): Phone survey, secondary data17 [159] Duan (2009): ecoinvent; China statistics18 [174] US EPA (2005) [Energy Star, certified]: Secondary data19 [164] Kawamoto (2001): Small-scale measurements20 [95] Hischier (2007) [ecoinvent]: Secondary data21 [177] MTP (2006): Measurements and secondary data22 [169] IVF (2007): Meta-analysis of prior studies23 [169] IVF (2007): Meta-analysis of prior studies24 [178] Schlomann (2005): Measurements and secondary data25 [161] Atlantic Consulting (1998): Secondary data26 [179] Nordman (2004): Estimates based on prior studies27 [178] Schlomann (2005): Measurements and secondary data28 [167] Choi (2006): Korea statistics29 [165] Hikwama (2005): Secondary data30 [13] Williams (2004): Secondary data31 [167] Choi (2006): Korea statistics32 [164] Kawamoto (2001): Small-scale measurementsFigure 3.4: Review of unit energy consumption of desktop PCsThe visual presentation of the data is illuminating; most studies are clustered between 100 and 350kWh/year, with an approximate average of 225 kWh/year. The highest data points are case studiesof a high-performance office workstation and a high-performance home gaming PC, highlighting thewide range of possible direct measurement results and thus the need for a large n if a study is to berepresentative of a typical PC. The lower outliers all tended to make questionable assumptions regardingpower consumption; for example, Kawamoto et al. [164] assumed that residential PCs are off 91% of thetime, but plug-load measurements by Porter et al. [176] show this figure to be closer to 60%. Likewise,Williams [13] assumed 3h of daily use for a residential computer, but Porter et al. [176] measured almost8h of daily use. Note that studies which report dominant impacts from the production phase rather than51the use phase - Williams [13] and Choi [167] - both report use-phase consumption at the low end of thisrange, suggesting that differing assumptions are largely behind the disagreement. Consumption of 50 to100 kWh/year as these studies assumed is not implausible for low-power desktops or for computers thatare used infrequently, but seems to be well below the representative range of a typical desktop undertypical usage patterns.3.5.2 Product lifespanThe lifespan of the product, multiplied by its annual energy consumption, determines the total energyconsumed in the use phase of a product. Measuring average product lifespan is deceptively difficult, be-cause computers can sit in storage for years after their useful life is over, and sometimes enjoy continueduse in the secondary market. The latter activity should be included in the total lifespan, as the product isstill consuming energy, but the former should not. Techniques to measure lifespan, which may includecustomer surveys, waste stream monitoring, or purchase monitoring, may not always be able to identifystorage and re-use, leading to variation in lifespan estimates.Studies measuring lifespan, and the estimates used in LCA studies, are shown in Figure 3.5. Theshape of the markers indicate whether the study was measuring first life only, or included re-use, ordid not specify whether or not re-use was included. The most persuasive study of lifespan to date isfrom Babbitt et al. [180], which calculated product lifespan based on 20 years of procurement data ina university setting (n > 2000 per year), and documented a steadily declining trend in lifespan; the lastreliable data was for purchases made in 2000, when the average lifespan was 5.5 years. Other reportedresults came from a survey of consumer purchases in Japan [181], a Gartner survey [182], and a seriesof studies in Japan [183, 184, 185], originally cited in studies by Yoshida [186, 187]1. Studies havebeen assigned a data quality rating, with larger primary studies ranking higher than secondary data, anddirect measurements such as through a procurement database as in [180] ranking higher than surveys.Notably, several life cycle assessments rely on unreferenced assumptions.As shown in Figure 3.5, estimates for PC lifespan range from three years to more than eight. Thebest-quality study, from Babbit et al [180], is probably close to an upper bound on first lifespan becauseit tracked computers on an academic campus with significant internal re-use. Downward trends in thatdataset suggest an average lifespan of 5 years in the present day. The smaller result from Gartner [182],1The original papers, in Japanese, were not available to us, so we could not confirm these figures.521990 1995 2000 2005 20100123456789Desktop PC lifespan (years)1234567891011 121314 1516171821Data QualityLowMediumHighPrimary dataSecondary dataIncludes reuseFirst life onlyUnknown/not specied1 [183] JEITA (2006) (Home): Japan study, cited in [187]2 [180] Babbitt et al. (2009) (1990 purchases): Procurement database3 [160] Tekawa et al. (1997) (Office): Unreferenced assumption4 [169] IVF (2007): Survey of computer vendors, n = 65 [184] Oguchi et al. (2006): Japan study, cited in [187]6 [159] Duan et al. (2009): Unreferenced assumption7 [166] Kemna (2005): Consumer survey8 [180] Babbitt et al. (2009) (1995 purchases): Procurement database9 [183] JEITA (2006) (Office): Japan study, cited in [187]10 [180] Babbitt et al. (2009) (2000 purchases): Procurement database11 [160] Tekawa et al. (1997) (Home): Assumption12 [165] Hikwama (2005): Assumption based on depreciation13 [185] ESRI (2007): Japan study, cited in [186]14 [168] Braune and Held (2006) [EPIC-ICT]: US EPA data15 [95] Hischier et al. (2007) [ecoinvent]: Secondary data16 [167] Choi et al. (2006): From [188]17 [162] Masanet et al. (2005): Aassumption18 [13] Williams (2004): From Williams and Hatanaka (2005) [181]19 [161] Atlantic Consulting (1998): Assumption based on depreciation20 [182] Smulders (2001): Gartner survey21 [181] Williams and Hatanaka (2005): Home survey, JapanFigure 3.5: Review of lifespan of desktop PCsa 3 year lifespan, was conducted in a business context which likely featured higher replacement frequen-cies than an academic context. Second lives due to reuse will increase the average lifespan. Accordingto Yoshida et al [186], the establishment of take-back recycling schemes increased re-use rates suchthat in 2004, about 37% of discarded PCs in Japan were re-used domestically, and an additional 25%were exported, though re-use rates are likely to be lower in jurisdictions like the USA which do nothave well-established recycling programs. The prevalence of re-use and the length of any secondary lifeare additional sources of uncertainty; very little high-quality data assessing these factors is available.Given the available evidence, lifespans between three and six years seem to be reasonable estimates,depending on the context of use and the availability of infrastructure to support re-use.3.5.3 Total use phase impact: analysisThe impacts due to electricity consumption can be measured in MJ primary energy or kg CO2e by mul-tiplying kWh of electricity by the appropriate impact characterization factors. Not all studies actuallyreport the emissions factors used, but in some cases they can be derived by identifying reported totaluse-phase impact, in MJ energy or kg CO2e, and dividing by total reported lifespan electricity consump-tion in kWh. These impact factors are shown in Figure 3.6, alongside emissions factors for various53electricity grids from the ecoinvent database [189], obtained using low-voltage at-grid supply data withcumulative energy demand (CED) and IPCC GWP100 LCIA schemes.Figure 3.6: Impact factors for electricity from LCA studies and ecoinvent databaseMost studies assume 1 kWh electricity is equivalent to between 10 and 12 MJ of primary energy inorder to account for losses in electricity generation. Williams’ study is an exception, accounting only fordirect electricity consumption in the use phase (1 kWh electricity = 3.6 MJ primary energy) [13], whichis a questionable assumption. Likewise, emissions factors are usually between 0.4 and 0.6 kg CO2e perkWh. Both of these figures are intended to encapsulate the impacts of the electricity generation infras-tructure at whatever geographic locale is appropriate for the study, so some variation is to be expected.In order to remove this variation from results in this study, constant factors of 11 MJ/kWh and 0.5kg CO2e/kWh are applied, which are the averages of the reported factors, excluding those of Williams.Note that these are slightly lower than most impact factors in ecoinvent; unfortunately, rationale for54impact factor choice is usually not presented in these studies, so it is difficult to determine why theywere chosen. The large spread in country data shows that the geographic location of product use is animportant determinant of the total impact of the product; for example, global warming impacts due touse phase will be thirty times higher in China than in Norway, as China’s electricity system is coal-based and Norway’s relies almost exclusively on hydropower. This information is not unknown to LCApractitioners, of course, but is not always clearly communicated in the presentation of results.The data in the previous sections suggest that plausible estimates for UEC range from 100 to 350kWh/year, and lifespan ranges from 3 to 6 years. Multiplying UEC and lifespan together gives a possiblerange from 300 kWh to 2100 kWh. Converted to primary energy using the average emissions factorsreported by the studies, this yields a range from 3300 MJ to 23100 MJ; in global warming potential,impacts range from 150 to 1050 kg CO2e. Absolute rather than probabilistic bounds are defined becausethe ranges are defined by qualitative assessment of the data rather than rigorous statistical aggregation,the latter approach being inappropriate due to the heterogeneity of the data. Consequently, the rangesare relatively large. The implications of this result, and the result of the analysis for the manufacturingphase, are examined in the next section.3.6 Analysis of overall impactFigure 3.7 shows the ranges identified in the previous two sections and their overall impact in terms ofprimary energy demand and global warming potential; also plotted are the results from other availablestudies. In interpreting these data, the weight of evidence suggests that the use phase impacts of atypical desktop PC are more likely to occur at the middle of the range, seen in Figure 3.7, and lesslikely to occur at the fringes. Studies that lie towards the bottom of the range, including Williams [13],Atlantic Consulting [161], Masanet [162], and Apple [170], have therefore calculated use-phase impactswhich are likely well below those of a typical desktop PC, though the Apple study examined a compactMac Mini system. The other Apple study [171] is well above typical, and examined a high-poweredworkstation (the Mac Pro); neither the Mac Mini nor Mac Pro are intended to be representative of atypical desktop PC. It is unfortunate that no methodological details are available for these studies asthis prevents any evaluation of their quality or reasonableness. Nevertheless, the very wide variation inoverall impact for these two studies is intriguing; if the methods are at least internally consistent, theninternal variation within the product category of desktop PCs may itself be large enough to overwhelm55any methodological or impact-data-related variation.05000100001500020000250003000011 1333444555666777Manufacturing Use TotalPrimary energy (MJ)050010001500200011 12225556 66777888999Manufacturing Use TotalGWP (kg CO 2 e)1: Atlantic Consulting 19982: Tekawa 1998 (includes CRT)3: Williams 20044: Kenma 2005 [MEEUP]5: Masanet et al 20056: Hischier et al 2007    [ecoinvent]7: IVF 20078: Apple 2010a (Mac Mini)9: Apple 2010b (Mac Pro)Figure 3.7: Overall primary energy and global warming potential, showing prior results and thisstudy’s estimates of reasonable ranges for a typical desktop PC without displayThe emissions factors for use-phase electricity consumption used in this figure are slightly belowthe average of the impact factors of European UCTE country electricity grids according to the ecoinventdatabase. Use-phase impacts will proportionally increase or decrease according to the impact charac-teristics of the local grid where consumption occurs. In very low-carbon jurisdictions like Norway,manufacturing impacts (which occur in southeast Asia) will far outweigh use-phase impacts.The collective evidence for impacts due to manufacturing is less conclusive. This study’s analy-sis identified significant variation in impact factors of subcomponents, especially mainboards and ICs,which were a substantial source of variation in the overall impact due to manufacturing, but there wasnot enough evidence to identify a reasonable range for these impact factors. Methodological differencesadd more uncertainty: Masanet [162] used a top-down input output method which is not easily compara-ble to those studies previously analyzed; the Apple studies [170, 171] did not report any methodologicaldetails. The complexity of the underlying methods and data makes independent verification of any ofthese studies very difficult. Reasonable ranges of manufacturing impacts are defined to span the rangedefined by those studies analyzed in Section 3.4, on the basis that they provide the most detail, mak-56ing an assumption that detail is correlated with quality and accuracy. Outlier studies could expand therange of reasonable results only if they are comparable in observable quality. This line of reasoningis admittedly somewhat speculative, but until standardization efforts reduce or eliminate uncertainty inproduct inventories and impact factors, accepting all high-quality studies as reasonable seems to be theonly defensible option.3.7 ConclusionWhen measuring primary energy consumption and global warming potential, the impacts due to usephase energy consumption of a typical desktop PC without display are the dominant impact in theproduct life-cycle, except in areas with very low-impact electricity grids; manufacturing impacts aresmaller, but still significant, and distribution and end-of-life are both negligible (note, however, thatend-of-life activities can cause significant damage to environment and human health not captured inthese two impact dimensions). A few studies have reported that manufacturing impacts exceed use-phase impacts, but these studies each reported use-phase energy consumption at or below the low endof a reasonable assessment of a typical PC’s energy consumption, which ranges from about 100 to 350kWh/year for a lifetime of 3 to 6 years. Manufacturing impacts exhibit a particularly high variability,most of which is due to disagreement about the impacts per unit of the various components inside the PC,especially mainboards and semiconductors. Estimates of the physical contents of a PC by mass vary aswell, but less so. Total impact, summarized in Figure 3.7, might be expected to range from 270 to 1300kg CO2e and from 4900 to 39100 MJ of primary energy, with the most likely result towards the middleof these ranges. The use-phase primary energy consumption and greenhouse has emissions are stronglydependent on the local electricity grid where the product is used. In addition, variation of the functionalunit, such as choosing a lightweight low-power computer or a high-performance workstation, can havea dominating effect on the value of the impact, making it important for researchers to either conductstudies with a larger sample of products in order to smooth out this variation, or to explicitly limitthe scope of their assessments to a specific sub-category. The latter approach has become increasinglyimportant with new products being introduced and the desktop PC category broadening.This exercise uncovered several inaccurate assumptions in published studies and highlights a gen-eral problem with life cycle assessments: they are very difficult to evaluate, even for an experiencedpractitioner, as a full listing of data, methods, and assumptions used is rarely available, often for reasons57of industrial confidentiality or proprietary data, and the correctness of the data may itself be difficult todetermine. A tendency for data re-use is noted as newer studies build on older studies, but as the pool ofprimary research is small and difficult to verify, errors may have been propagating through the literatureundetected, creating a risk that erroneous results may become established. Data on electronics manufac-turing, especially for less-frequently researched components and processes, is particularly vulnerable toundetected errors, and current reported results are of an unknown quality.Data quality issues will diminish when standardized impact-reporting systems with participationfrom electronics manufacturers are in place. Until such time, practitioners should use caution whenadapting results from previous studies and critically evaluate such results regardless of their prominenceand apparent acceptance by other researchers. Rich opportunities remain for future research to reduceuncertainty in life cycle assessments of electronics.58Chapter 4Embodied emissions of ICT devices4.1 IntroductionThe global share of worldwide greenhouse gas (GHG) emissions from information and communica-tions technology (ICT) is substantial and rising; computers and electronics are a significant source ofhousehold electricity consumption [190]. In the personal computing (PC) sector, operational impactsare estimated to account for roughly 60% of greenhouse gas emissions, with the remaining 40% due tomanufacturing [25]. The latter, also referred to as embodied emissions, is difficult to estimate, and thereis a need for both additional and improved estimates of embodied emissions of ICT products, and forheuristic methods to enable faster and easier first-order estimation.Current literature examining the embodied impacts of ICT equipment suffers from three importantshortcomings: disagreement across studies regarding the magnitude of impacts of ICT products [1,19, 20]; lack of coverage for newer products; and lack of transparency in studies, particularly due toconfidential input data, which hampers reproducibility and cross-study comparisons. Using primarydata from hand disassembly and the ecoinvent v.2.2 database [191] for upstream process data, this studyquantifies the embodied greenhouse gas (GHG) emissions of 11 ICT products, most of which weremanufactured in 2009 or later. This work represents the first peer-reviewed examination of the embodiedimpacts of a small-form-factor desktop PC, netbook-style laptop, thin client device, Apple iPad, AppleiPod Touch, and Amazon Kindle. An additional study of three older ICT products from the ecoinventdatabase [95] was reformulated using this study’s framework so that all 14 products could be compared.A full listing of the products analyzed and the results recorded is in Table 4.1. By using a consistent59framework, product emissions estimates can be compared against one another with some confidence,avoiding the problems of different modeling assumptions or different upstream data sources that arisewhen comparing results from independent studies. This framework is used to develop first-order linearmodels for estimating embodied emissions using a small set of product characteristics. Findings arealso compared against a dataset published by Apple (described in Appendix B) that provides life cycleassessments (LCAs) of its entire product line [46].Product category Model (year of manufacture) Mass [kg] GHG [kg CO2e]Desktop PC [ei] Typical desktop (2002) 11.1 322Dell Optiplex 780 mini tower (2010) 10.7 164Dell Optiplex 780 ultra-small form factor (2010) 3 73.5Dell FX-100 zero client (2009) 1.3 33.6Laptop PC [ei] 12.1” HP Omnibook with dock (2003) 3.3 256HP 530 laptop, 16” (2009) 2.8 108HP Mini 110-1030 CA Netbook, 10” (2009) 1.3 62.2LCD display [ei] Typical 17” (2004) 5.1 297Samsung Syncmaster 2243 21” (2009) 5.1 168Mobile electronics Apple iPad 8 GB Wi-Fi 1st gen (2009) 0.78 25.5Apple iPod Touch 8 GB 3rd gen (2009) 0.2 7.5Amazon Kindle Wi-Fi 3rd gen (2010) 0.31 13.3Server and network Dell PowerEdge EMU3710P71 rack server (2005) 15.5 3833Com 24-port Superstack 3 10/100 Ethernet switch (2003) 2.1 91.8[ei]: Adjusted version of study originally published in ecoinvent [95]Mass includes power supplies; desktop PCs exclude displayGHG: embodied emissions derived in this studyTable 4.1: Products analyzed in this studyThe broad goal of this work is to make LCA results for ICT products easier to derive and more usefulin supporting decisions, both by contributing a new primary dataset of product inventories and impactestimates, and by exploring linear regression-based models that could approximate impact assessmentusing a limited set of easily collected inputs. Similar linear regression-based methods being developedby the PAIA project [192] and by iNEMI [193] are aimed at enabling impact estimation using productcharacteristics (e.g. screen area, amount of RAM, hard drive size, etc.). The dataset produced by thisstudy could be adapted to use these tools and methods once they become publicly available.604.2 MethodologyThe process-LCA methods applied here have been used in many studies of ICT equipment, such as theadapted ecoinvent studies [95], the EPIC-ICT project [168], the EU energy-using-product studies [166,169], and others. The ecoinvent database has been described as the “most complete and transparentprocess-LCA database” [81] and is used for upstream data in one LCA study of a desktop PC [159]and as the process component in several hybrid-LCA studies [92, 138]. However, the limitations ofthis methodology must be stressed. Process-LCA accounts for only those impacts that are specified inproduct inventories and underlying process databases; truncation error due to activities not modeled inthese databases can be significant. Large sectors of the economy, especially service sectors, are notmodeled by the ecoinvent database at all [81]. Top-down methods, such as economic input-output LCA,do not suffer from truncation error, but have a limited ability to distinguish between similar productsdue to the coarseness of economic data. Hybrid-LCA methods attempt to achieve a balance by mergingtop-down economic data with process-LCA results. Two hybrid-LCA studies of a desktop PC [13] andlaptop PC [92] found that the economic correction respectively accounts for 51% and 40% to 56% oftotal impacts in the production phase, respectively. Likewise, in a comparison of LCA methods for ICTproducts [138], the original process-LCA estimate accounted for only 37% of the emissions estimatedby a top-down input-output LCA. Accordingly, hybrid analysis is recommended by several researchersas the best means to produce accurate estimates of emissions in absolute terms [81, 138, 139].The scope of this study is limited to comparing the embodied impacts that can be identified usingprocess-LCA methods with the ecoinvent database. This limitation is imposed because the strengths andweaknesses of this framework are relatively well understood, which allows for increased confidence thatthe relative differences in the product impacts that the analysis identifies are not methodological arti-facts. The process-sum method introduces significant truncation error such that this study’s results un-derestimate the absolute impact; the use of economic data to correct for truncation error would improveaccuracy, but such a correction would require pricing data which are largely unavailable. Accordingly,these results should be interpreted as a comparison between products, and not as a calculation of productcarbon footprints that could be used in contexts external to this study. Future work to produce productcarbon footprints should address this truncation error, as well as the use phase and end-of-life phasewhich are not included in this study.61A number of standards for LCA of ICT equipment have recently been published or proposed bythe ETSI [194], ITU [195], IEC [196], and GHG Protocol [197], coordinated in part by the EuropeanCommission’s Information Society department [198]. This study was conducted prior to the publicationof these standards, and thus is not compliant with any of them, though compliance could be achievedwith some additional effort. The use of standards, especially the ETSI standard which is the most thor-ough and rigorous of the above, would greatly improve transparency and comparability across compliantstudies, and should be encouraged.This study focuses on embodied impacts, in terms of global warming potential, using the IPCC100-year characterization; primary energy demand is calculated as well in Appendix B. Raw materialextraction, processing, final assembly, and transport are included. Modeling assumptions are equivalentto those used in previous ecoinvent studies [95]. In particular, the ecoinvent database supplies data forall electronic component manufacturing, processing, assembly, and transport data and assumptions, withthe exception of silicon die, for which an updated study is used [172].Bills of materials were constructed via hand disassembly and weighing. Product packaging and extraparts, including manuals, software, and extra cables and adapters, are excluded, as this information isnot available for all products and tends to represent a small share of the impacts for ICT products. Someother components, such as flame retardant coatings, are not detectable via weighing and were thusexcluded. The ecoinvent models for silicon die content in packaged integrated circuits (ICs) containa calculation error (identified in other studies [1, 133]) that leads to an underestimate of silicon diecontent. In order to develop more accurate estimates of silicon die content in packaged ICs, a selectionof packaged ICs were X-rayed and their die areas measured. Silicon die content is modeled as a linearfunction of packaged IC area for large ICs and of mass for smaller ICs, which were weighed in bulk.An additional correction was made for stacked ICs in the Apple iPad and iPod Touch.Appendix B contains full details regarding modeling assumptions; the mapping of bills of materialsentries to ecoinvent processes; experimental data and calculations for the silicon die ratios; and a dis-cussion of stacked ICs. Full bills of materials for all products analyzed are provided in the SupportingInformation of the published study available online [2], but omitted from this dissertation due to theirlength.624.3 ResultsTwo sets of results are presented: the mass composition of each product as determined through handdisassembly (Figure 4.1a), and estimated embodied GHG emissions calculated with ecoinvent impactdata (Figure 4.1b). Data tables for the graphs are available in Appendix B.4.3.1 Product composition by massCircuit boards account for between 5% and 20% of product mass across most products. Casing, typicallymetal or plastic, represents roughly half of the mass in large desktop and rack servers, each of whichweigh more than 10 kg. In mobile devices, casing is only about one quarter of the mass due to the extramass of batteries and displays.The three studies adapted from the ecoinvent database model a desktop PC, manufactured in 2002;a 17” LCD monitor from 2004; and a 12.1” laptop (with dock) from 2003. Comparable products inthis study, all manufactured in 2009 or 2010, are significantly lighter. The 21.5” monitor studied hereis comparable in mass to the 17” monitor modeled in ecoinvent, despite the former’s larger screensize, because the latter was significantly thicker and had a much heavier frame, likely because LCDtechnology was relatively new and less compact in 2004. This information constitutes evidence thatelectronics products are becoming more materially efficient over time.In terms of both IC mass and die area, modern devices show significantly less material usage forintegrated circuits when compared to the older products modeled in ecoinvent. That is likely due tohigher levels of miniaturization available in modern packaging technologies (as described in anotherstudy [93]), as well as reduced numbers of ICs per product due to increased integration of functionality.4.3.2 Embodied emissionsCircuit boards including ICs are responsible for the majority of embodied GHG emissions in most de-vices. Product casing in this study is modeled as aluminum, steel, or plastic, all of which have relativelylow emissions per unit mass, so the overall impact of casing is small. One exception is the laptopmodeled in the ecoinvent database for which the casing includes a higher-emissions magnesium alloy.Integrated circuits have high impacts despite their very small mass; silicon dies alone are responsiblefor about 20% of product embodied emissions on average.The results suggest a strong link between product mass and embodied emissions, with heavier or63a) MassMeasurement (g)050001000015000OtherDisplayBatteryPower supplyCasingIC's (die)IC's (packages)Circuit boards (excl. IC's)Percent of total (%)020406080100Desktop (ei)Desktop − towerDesktop − smallThin ClientLCD monitor, 17" (ei)LCD monitor, 21.5" Laptop, with dock, 12" (ei)Laptop, 16"Netbook, 10"iPadiPod touchKindleRack serverSwitchb) Embodied greenhouse gas emissionsEstimate (kg CO2 e)050150250350OtherAssemblyTransportDisplayBatteryPower supplyCasingIC's (die)IC's (packages)Circuit boards (excl. IC's)Percent of total (%)020406080100Desktop (ei)Desktop − towerDesktop − smallThin ClientLCD monitor, 17" (ei)LCD monitor, 21.5" Laptop, with dock, 12" (ei)Laptop, 16"Netbook, 10"iPadiPod touchKindleRack serverSwitchFigure 4.1: Results showing product mass (a) and embodied emissions (b). (ei) denotes adjustedstudies from ecoinvent database.64larger products having higher emissions; the following section explores this relationship in more detail.The older devices modeled in ecoinvent (desktop, laptop, monitor) have significantly higher emissionsthan modern devices with similar functionality. Since the modeling framework and data sources areidentical in this study and in the ecoinvent study, aside from some minor adjustments discussed inAppendix B, the differences can be ascribed to changes in the material composition of the productsthemselves: modern devices have fewer integrated circuits and circuit boards, certainly a consequenceof higher levels of on-chip integration enabled by Moore’s Law. When modern devices are compared toother modern devices from the same product category, smaller form factors have smaller impacts: the10” netbook’s embodied emissions are about 40% lower than those of a similar 16” laptop, while thesmall desktop’s emissions are about 50% lower than a comparable minitower from the same productline. The embodied GHG emissions of mobile devices are very small compared to laptops, desktops,and monitors.4.3.3 Comparison with other studiesThis study’s output estimates for the adjusted ecoinvent studies are very similar to those from the origi-nal ecoinvent studies, suggesting that the framework has been accurately reproduced; differences due tothis study’s adjustments are explained in Appendix B. Compared to recent studies of similar equipmentpublished within the last five years, this study’s results are mostly similar. This study’s emissions esti-mate for a rack server, 360 kg CO2e, is comparable to a recent study’s estimate of 380 kg CO2e [133],while this study’s result for the Dell Optiplex 780 tower desktop, 161 kg CO2e, is comparable to Dell’sestimate of 120 to 180 kg CO2e for the same product [199]. Dell’s study of a 14” laptop estimatesemissions to be 160 kg CO2e [200], larger than this study’s estimate of 106 kg for a 16” HP laptop. Inthis case, differences occur primarily in casing (25 kg CO2e in Dell’s study vs. 6 kg CO2e in this study),mainboard (72 kg CO2e vs. 52 kg CO2e), and battery (9 kg CO2e vs. 1.3 kg CO2e). The latter studyis the only example from a manufacturer in which emissions are specified at a component level thatallows a detailed comparison. A recent hybrid-LCA study of a 15” laptop manufactured in 2001 foundGHG emissions from manufacturing to be between 227 and 270 kg CO2e; of that amount, 93 to 136kg CO2e were accounted for via bottom-up process LCA, comparable to our estimate of 106 kg CO2efor a 16” laptop using similar methods, with the remaining 134 kg CO2e due to the top-down economiccorrection.65There is one study of an e-book reader, and it identifies 40 kg CO2e for a 2007 Sony PRS 505 [201],higher than this study’s estimate of 13 kg CO2e for a 2010 Amazon Kindle. That study also uses theecoinvent dataset, but identifies 31 g of packaged ICs compared to 2 g in this study, implying thatthe difference is due to physical variation between products. Two studies of mobile phones report 20kg CO2e [202] and 30 kg CO2e [203], higher than this study’s estimate of 7 kg CO2e for an iPod Touch.The mobile phone in the latter study had a mass of 250 g, which may or may not include an externalcharger, whereas the iPod Touch has a mass of 109 g, while its charger’s mass is an additional 89 g.Given the tendency for the newer products in this study to use fewer integrated circuits compared toolder products in the ecoinvent database, we speculate that the older mobile phones in those two studieslikely contained more ICs, which could account for some of the difference between these results.Apple’s dataset of 22 products [46]1 shows results that are considerably higher than this study’sestimates for similar products. This study’s estimates of the manufacturing emissions of a laptop, net-book, iPad, and iPod touch are 106, 62, 22, and 6.7 kg CO2e, respectively; comparable products in theApple product environmental reports, a 15” Macbook Pro, 11” Macbook Air, iPad 2, and iPod touch,are estimated to have embodied emissions of 290, 162, 25, and 15 kg CO2e, respectively. The author ofApple’s reports indicates that a different impact database, i.e. not ecoinvent, was used by Apple, andnotes that the casing materials were modeled in more detail and were not classified simply as aluminumor steel, but as rather more complex materials, though the methodologies are otherwise comparable2. Alack of transparency and access to the details of Apple’s study prevents definitive identification of thesource of the variation.Overall, when compared against other studies, this study’s results are roughly comparable for largeproducts and lower for smaller mobile products. In some cases, the variation may be a consequence ofthe more recent vintage of products analyzed here relative to other studies, since newer products tendto have fewer integrated circuits. Additional variation may be caused by different underlying modelsfor some parts (such as casing and battery) and/or different modeling assumptions. It must be stressedthat this study is replicating the modeling assumptions of the ecoinvent database. Agreement betweenthis study’s results and those of other studies is not a guarantee of accuracy, which is a function of theunderlying source data and methods. Such agreement does, however, suggest that this study’s methods1This study considers a historical snapshot of Apple’s product environmental reports as of September 2011, when the studywas conducted; environmental reports released or updated after that time are not accounted for.2Personal communication, 201166and data are in-line with standard practices. To the extent that these practices are valid, the relativedifferences in our estimates of embodied emissions for different products arise due to differences in theproducts being analyzed, as intended.4.3.4 Data quality and uncertaintyThe ecoinvent database uses a semi-quantitative uncertainty model based on “pedigree” matrices inwhich each inventory item is assigned a probability distribution based entirely on the quality of thedata as estimated by experts [191]. These scores are intended to account for mismatches between theecoinvent technological processes and the physical real-world processes they model. The same methodis applied here in order to enable comparisons with other studies that use the ecoinvent database anduncertainty method, using Monte Carlo analysis with 100 trials per product. Details about the probabil-ity distributions used are discussed in Appendix B; results are shown in Figure 4.2. The distributionshave standard deviations ranging from 10% (several products) to 18% (LCD monitor) of their respectivemeans.This uncertainty model has some advantages in that it is quantitative, consistent, and tractable,but it relies on expert judgment and is prone to errors related to that approach [204]. In particular, thepedigree approach produces artificial probability distributions that have no empirical basis and representonly expert judgments of data quality, and in addition only capture uncertainty which is internal tothe ecoinvent modeling framework. Structural uncertainties due to truncation errors are not included;neither are uncertainties in emissions characterization factors. The actual bounds on the results will belarger than those shown in Figure 4.2 and are not precisely quantifiable using this method.4.4 AnalysisThe data in Figure 4.1 suggest the presence of a linear relationship between embodied emissions andproduct mass. This trend also appears in Apple’s dataset (n=22). These datasets could be used toestimate embodied emissions for ICT products based on easily measurable physical characteristics suchas total mass and volume. Differences in the underlying modeling frameworks mean that this study’sdataset and Apple’s cannot be combined in such an analysis. However, it is possible to develop a linearmodel that adequately describes both datasets, tuning the coefficients independently to produce one setof coefficients for each dataset.670100200300400Rack serverDesktop (ei)LCD monitor, 17" (ei)Laptop, with dock, 12" (ei)LCD monitor, 21.5"Desktop − towerLaptop, 16"SwitchDesktop − smallNetbook, 10"Thin clientiPadKindleiPod touchEmbodied GHG emissions (kg CO 2 −eq)Figure 4.2: Monte Carlo results: mean embodied GHG emissions with error bars showing ± twostandard deviations, using data quality pedigree matrix approachApple provides a breakdown of the material composition of its products by mass. By arranging thedata from this study’s disassembly analysis into the same categories as Apple, six possible predictorsof embodied GHG emissions are yielded: circuit board mass, display mass, battery mass, casing mass,power supply mass, and other mass. Two additional predictors that describe overall product characteris-tics are included: product mass and product volume. Using combinations of these eight predictors, manypossible linear regression models were examined, each of which was independently fit to both datasetsto produce a set of coefficients that could be used to predict a product’s emissions. Model robustnesswas evaluated using leave-one-out cross-validation in order to obtain the cross-validated residual sum ofsquares statistic (cvss), which is intended to quantify how well the model can predict output values fordata points not in the training set. Each candidate model is assigned two scores, cvss1 and cvss2, whichare equal to the cvss for this study’s dataset and for Apple’s dataset, respectively. A combined score forboth models is generated by summing cvss21 + cvss22; the sum of squares is used to penalize models thatfit one dataset well and the other poorly, since the goal is to find a model that fits both datasets well. The68linear model that has mass as its only predictor is used as a benchmark against which other scores arenormalized.Three of these models are shown below in detail; more model results are in Appendix B. The first,‘mass only’, which is the benchmark model, treats emissions as a linear function of product mass. Thesecond model, ‘pcb+disp+batt’, uses three internal predictors (i.e. Emissions = α1 ·masscircuit board +α2 ·massdisplay +α3 ·massbattery). The third, ‘all internal’, uses all six internal predictors (circuit board,display, battery, casing, power supply, and other). All models are constrained to pass through the origin;adding an intercept does not improve the fit. For each model, the fitted results for both datasets areshown in Figure 4.3 and the coefficients and diagnostics in Table 4.2. In the figure, the x-axis representsthe actual estimate for embodied emissions for each product in the dataset, while the y-axis shows theresidual; a perfect model would have each point along the y = 0 line.This study Apple data−200−1000100−200−1000100−200−1000100llll lllllllllllllllllllllllllllllllllll llllllllll lllllllllllllllllllllllllllllll lllllll llllllll pcb+disp+batt Mass only All internal0 200 400 600 800 0 200 400 600 800Product embodied emissions (kg CO2 −eq)Residual (kg CO 2 −eq)Figure 4.3: Residuals from model selectionThe ‘mass-only’ model has the best cross-validated score of the dozens of models that were com-pared, while the ‘pcb+disp+batt’ model is only slightly worse. Figure 4.3 shows that the mass-onlymodel systematically underestimates emissions for light products, especially in the Apple dataset; that69Model coefficient [kg CO2e per g] p-value‘pcb+disp+batt’ model ‘mass only’ model ‘all internal’ modelPredictor This study Apple data This study Apple data This study Apple dataMass 0.027 <0.01 0.039 <0.01pcb 0.18 <0.01 0.37 <0.01 0.24 <0.01 0.48 <0.01casing 0.012 0.49 -0.1 0.23batt 0.30 <0.01 0.36 <0.01 0.36 0.02 0.41 <0.01disp 0.065 <0.01 0.052 <0.01 0.062 <0.01 0.066 <0.01psu -0.079 0.97 0.53 0.02Other -0.012 0.41 0.1 <0.01cvss 0.42 1.4 1.0 1.0 0.45 2.8Score 2.1 2.1 2.0 2.0 7.9 7.9R2 0.97 0.94 0.85 0.89 0.97 0.97Adj.R2 0.96 0.93 0.84 0.88 0.95 0.96Table 4.2: Coefficients and statistics from model fittingoccurs because the masses of heavier products are dominated by casing, which has a much loweremissions-per-unit-mass than electronic components. A linear-mass model is not sufficiently sophis-ticated to account for the different composition of light and heavy products. The ‘pcb+disp+batt’ modelappears to be unbiased, and has a comparable cross-validation score, meaning it is a better fit over-all, though it requires more data inputs than the mass-only model. The ‘all-internal’ model illustratesthe effect of including additional predictors: the residuals are smaller, but the model is overly tunedto the dataset and does not accurately predict emissions, as shown by its poor cross-validation score.Some coefficients for this model are negative, indicating non-physical results and double-counting fromcorrelated predictors.The two best models - ‘mass only’, and ‘pcb+disp+batt’ – both produce good results and coeffi-cients that are physically reasonable and fairly consistent numerically across the two datasets. Ideally,the analysis could be repeated with a wider dataset, perhaps including multiple specimens from eachproduct category, and should incorporate newer process data when available. Nevertheless, the strengthof the linear relationships uncovered in both this study’s dataset and Apple’s dataset is encouraging,and suggests that linear models based on a limited number of product characteristics could reasonablyapproximate the results of using full process-sum LCA to estimate manufacturing emissions.704.5 DiscussionThis work compares modern ICT products to those originally modeled almost a decade ago in theecoinvent database. In all three cases, the newer products’ embodied GHG emissions are an estimated50% to 60% lower than those of the corresponding older products. This decrease is mainly caused bya reduction in total mass and a proportional decrease in integrated circuits and circuit boards, the resultof systems becoming more highly integrated and thus using fewer ICs. These products were chosen tobe roughly representative of typical products within their respective categories, so it can reasonably beinferred that over time ICT products are getting lighter, becoming more integrated, and having a reducedimpact. However, efficiency trends in ICT products are counteracted by increased growth in the installedbase of existing products and the emergence of new complementary products; the overall impact of theICT industry, or of a consumer’s personal collection of ICT products, depends on the relative strengthof these competing trends, which are not analyzed here.Embodied impacts identified in this study are roughly linear with respect to mass, with a coefficientof 27 kg CO2e per kg of product using this study’s data and 39 kg CO2e per kg of product using Apple’sdata, but this model tends to underestimate impacts of lighter products. If the masses of printed circuitboard (including ICs), battery, and display can be obtained, then emissions can be calculated as 0.18 ·masspcb + 0.30 ·massbattery + 0.065 ·massdisplay (using this study’s data), or as 0.37 ·masspcb + 0.36 ·massbattery +0.052 ·massdisplay (using Apple’s data), with masses in grams and coefficients in kg CO2eper gram. These results do not account for truncation error and exclude the use phase and end-of-lifephase; they therefore should not be taken as product footprint benchmarks. More sophisticated linearmodels could be constructed, but the benefits of doing so appear small.The relatively good fit of these models suggests that linear regressions based on product characteris-tics may be a promising avenue of exploration for first-order life cycle assessments. Further investigationcould determine the stability of these findings over a wider range of product categories, as well as overvariation within a product category; for example, by assessing several monitors with different screensizes, as in the PAIA project [205], or by examining the effects of using alternative materials (e.g. forcasing) that may have considerably higher impacts per unit mass than the materials used here. In addi-tion, the effect of imposing a top-down economic correction on the results could be explored. Finally,the necessity to re-tune the model depending on the underlying framework and data is problematic; en-71forcing compliance with the aforementioned ETSI standard [194] (or others) could enable standardizedmodels for simplified LCA, which would help achieve the broad goal of a more practical and usefulmethodology to support decisions.72Chapter 5Emissions due to connected media5.1 IntroductionInformation and communications technologies (ICTs) represent a small but growing share of the globalenergy and CO2 burden, accounting for roughly 2 to 3% of global anthropogenic greenhouse gas (GHG)emissions and 4 to 6% of global electricity consumption [25]. Growing convergence between televi-sion sets and personal electronic devices, alongside a shift from traditional broadcast media distribu-tion forms to on-demand streaming over the internet, blurs the boundary between the formerly distinctcategories of ICT and consumer electronics (CE). When these two categories are grouped, they ac-count for about 12.5% of combined residential and commercial building energy consumption in theUS [26]; this total does not include most datacenters, which themselves account for about 2% of USelectricity consumption [23]. Overall energy consumption due to ICT/CE devices in the home anddue to network and datacenter infrastructure are well-known through prior studies, especially in a UScontext [5, 206, 207, 208]. However, connected media platforms such as the tablet, smartphone, andconnected TV account for a growing share of consumer time spent with digital media [98, 209] whichcomplicated the relationship between end-uses, such as watching video on a device, and energy con-sumed in the provision of that end-use.ICT/CE devices have many end-uses, such as watching video, consuming online content, and others.Greenhouse gas emissions due to such end-uses may be assessed with a life cycle assessment (LCA)approach, which tabulates all emissions associated with operation of electronic devices and infrastruc-ture which services the end-use, as well as all embodied emissions which occur during the production,73assembly, transportation, and disposal of all devices and infrastructure involved in servicing the end-use. Previous life cycle environmental impacts of individual ICT products have not included the energyor emissions due to services typically accessed on those devices. Such work has focused on desktopPCs [1], laptops PC [210], mobile phones [211], and other devices to a lesser degree such as televisionsand servers [2, 19, 91, 212]. Studies of the impacts of ICT-enabled behaviors have drawn wider bound-aries, such as comparing changes in travel impacts due e-commerce [130, 213] or telework/telecom-muting [44, 68, 214], or comparing the impacts of physical media against online textbooks [121], e-books [38], online journals [124], online news [118, 126], digital music [39], and streaming video [40].The studies have varying goals, but each tests the impact of a relatively new ICT-enabled behavior, andeach includes a comparison to a functionally similar non-ICT behavior. However, comparisons of phys-ical and digital end-uses are prone to large uncertainties since they entail comparisons of systems withvastly different characteristics.More recently, researchers have begun to examine the impacts of ICT/CE end-uses in absolute termsand in comparison with other ICT/CE end-uses. Study topics include broadcast TV compared withon-demand video [45]; impacts of video and text online news delivery on different platforms [41];comparisons of business productivity software with “cloud”-hosted alternatives [42, 129]; a comparisonof access network, core network, and datacenter energy for wireless cloud applications [43]; and aforecast of total global ICT/CE energy through 2017 [215]. In contrast to previous studies, these studiesdo not compare against non-ICT behaviors; rather they seek to better understand the relative impacts ofdifferent behaviors, often by taking a national or global approach; to identify environmentally preferableor non-preferable behaviors; and to identify the most impactful components of ICT/CE systems so thatfuture impact reduction efforts may be effectively targeted. This approach allows for the developmentof higher-quality models and data inputs given the more focused scope; in addition, it enables linkagesbetween well-established estimates of aggregate ICT/CE energy, and the behaviors that are driving it.This work continues along this approach by focusing on digital media end-uses in the home acrossfour prominent media platforms, in order to better understand how consumer digital media consumptionhabits relate to overall ICT/CE emissions. The remainder of this paper is organized as follows. Sec-tion 5.1.1 outlines the study boundaries and the scope of this work. Section 5.2 describes the modelingmethodology which is used to evaluate energy use and GHG emissions relating specific end-uses (e.g.video, browsing, etc) to device, network, and datacenter combinations. Section 5.3 describes model74inputs estimated from US data. Section 5.4 presents energy use and GHG emissions by device types(5.4.1), end uses (5.4.2), and presents an uncertainty analysis (5.4.3). The study also compares ICTuses in the home to those from other household appliances (5.4.4) and concludes by summarizing theimplications of this study’s findings for future research in Section Study scope and boundaryMetro/edge &core networksMobileaccess networksWi-Fi / fixedaccess networksTV platformPC platformTabletSmartphoneMedia platformsEnd usesBroadcast TVOn-demand TVOnline videoOther online usesOffline usesDatacentersFixed networkMobile networkTimespentDatatrafficDatatrafficEmbodied and operational greenhouse gas emissions: US consumers, 2012 and 2017Figure 5.1: Block diagram showing study scope and boundaryFigure 5.1 shows the study scope and boundary. The study estimates the embodied and operationalgreenhouse gas emissions of five broadly defined end-uses using four ICT/CE platforms among USconsumers in 2012 and 2017. The study applies a top-down methodology using secondary data toestablish emissions intensities (i.e. emissions per unit of data transfer) for networks and datacenters.Forecasts to 2017 derived based on best available secondary data sources were used to show the effects ofcontinuing growth in data traffic alongside continual improvement in network and datacenter efficiency,both of which are changing rapidly. The time horizon of 2017 was chosen to coincide with the horizon ofthe Cisco VNI data traffic forecast [103], a major data source for this study, as well as horizons of otherdata sources regarding energy consumption and installed base. A key contribution of this study is thedevelopment of an integrated end-use time and traffic model that allows for calculation of greenhousegas emissions of each end-use on each platform. The model uses a compilation of appropriate datainputs, relying primarily on data from market research and technology company reports; it is transparent75and could be easily applied to other contexts with different intensity estimates.Five end-uses are modeled: (1) broadcast television, i.e. TV programs delivered via aerial broad-cast or over cable or satellite networks; (2) IP video-on-demand, i.e. TV programs delivered to con-nected televisions over managed IP networks, usually through a service managed by a cable or satelliteprovider; (3) online video, also known as “over-the-top” video, which includes Youtube, Netflix, Huluand all other internet video services; (4) other online uses, such as online gaming, streaming music,and social network; and (5) all offline uses including word processing, DVD watching, offline gaming,and other uses which do not involve network data transfer. These categories of end-use were chosenso that they reasonably encompass the majority of ICT/CE behaviors. Four device platforms are con-sidered: TVs, PCs, tablets, and smartphones. Networks are modeled as a combination of fixed accessnetworks, which include customer premises equipment (e.g. routers) and customer access equipmentrun by telecommunication operators; mobile access networks, including cellular base station for all net-works in operation; and core IP networks, which includes all metro, edge, and core network equipment.A synthetic “fixed network” is defined to facilitate model calculations. It includes a fixed access networkand that portion of core IP networks which carry data traffic that arrived via fixed access networks; a“mobile network” is similarly defined. Datacenters include all servers and storage equipment as well ascooling, power distribution, and other overhead. This network model is simplified; the intent is to be asinclusive as possible when accounting for energy and emissions.5.2 MethodologyDevices that are used in conjunction with one another to perform end-uses are grouped together as“device platforms”. Tablets and smartphones are each platforms which consist of one device only – thetablet device and smartphone devices themselves. The TV platform includes a number of peripheraldevices: set-top boxes, game consoles, DVD/Blu-ray players, and A/V receivers, as well as the TVset itself. Because not all of these devices will be present in each case, an average TV platform isdefined which contains one TV set and a fraction of each peripheral according to the relative installedbase, with impacts scaled accordingly; for example, if there are 0.6 set-top boxes per TV set, then theemissions of using the TV platform include the emissions of the TV set plus 60% of the emissions ofone set-top-box. Other platforms defined in this study are the desktop PC platform, which includes LCDmonitors; the laptop PC platform, which also includes LCD monitors but in a smaller proportion; and76the average PC platform, which includes a mix of desktop PCs, laptop PCs, and LCD monitors. Otherperipherals are excluded on the basis that their aggregate energy consumption is relatively small [5].Platform definitions follow in Section 5.3.Greenhouse gas emissions are assessed using a life cycle assessment (LCA) approach with impactstabulated in terms of kg CO2e/yr. An end-use, EU , such as watching online video, is undertaken throughthe use of a platform P, which consists of one of more devices D. The emissions of performing end-useEU on platform P at time t are defined as GEU,P(t). Total US emissions are determined by the platforminstalled base, as follows:GEU,P,total(t)[kg CO2eyr]= IP[platforms] ·GEU,P(t)[kg CO2eplatform ·yr](5.1)Emissions GEU,P(t) are the sum of emissions due to devices, networks, and datacenters. Each isdiscussed in turn below.5.2.1 Device emissionsFor a given device, D, emissions due to end-use EU are specified as GEU,D. Emissions at the device levelare allocated according to the amount of time spent pursuing the end-use on the associated platform,TEU,P, relative to the total amount of time spent actively using the platform for all end-uses, Ttotal,P. Thetime share given by TEU,P/Ttotal,P is applied to annual device emissions, as follows:GEU,D(t) =TEU,PTtotal,P(GD,emb.(t)LD(t)[kg CO2edevice ·yr]+ED(t)[kWhdevice ·yr]·EF[kg CO2ekWh])(5.2)Device emissions are calculated as the sum of annualized embodied emissions, equal to overallembodied emissions per device GD,emb. divided by average device lifespan LD, plus annual operationalemissions, equal to device annual operational energy use ED(t) multiplied by a grid emissions factorEF . Throughout this analysis, a US average grid emissions factor of 0.6 kg CO2e/kWh is applied[216]. The rightmost factor thus represents the annualized GHG footprint of the device, accounting forembodied and operational emissions, while the left factor is a percentage share allocated according totime spent. The two factors are related through the total amount of time spent which influences overallpower consumption, but the annualized footprint is defined to include energy consumed while the device77is idle, while Ttotal,P includes only time spent pursuing an active end-use. As such, overhead emissionsare proportionally allocated to active end uses, on the basis that the device was purchased and turned onin order to pursue these end-uses. This is done to ensure that all emissions are accounted for.The emissions from all of the devices D which are part of platform P are combined as follows:GEU,P,devices(t) = ∑D∈PIDIPGEU,D(t) (5.3)Here ID is the installed base of the device, and IP is the installed base of the platform, so that emissionsdue to devices which are part of platforms are scaled according to their prevalence. In other words,a platform is made up of a weighted sum of devices with the weights determined by relative installedbase.5.2.2 Network emissionsTwo types of networks are modeled: fixed, and mobile. GEU,P,NF is the emissions due to fixed networkinfrastructure, including fixed access networks and core IP networks, due to performing the end use onone instance of platform P; likewise GEU,P,NM is the emissions due to mobile network infrastructure,including mobile access networks and core IP networks. Network emissions are allocated according tothe amount of data traffic generated by the end-use. The total amount of traffic generated by end-useEU on platform P is defined as DEU,P, measured in gigabytes (GB) per year, which is divided into datatraffic on fixed and mobile networks:DEU,P[GBplatform ·yr]= DEU,P,NF +DEU,P,NM (5.4)Assuming fixed networks carry total data traffic of DNF GB/yr, emissions due to fixed networkinfrastructure are calculated as follows:78GEU,P,NF (t) =DEU,P,NF (t)DNF (t)(GNF ,emb.(t)[kg CO2eyr]+GNF ,op.(t)[kg CO2eyr])(5.5)=DEU,P,NF (t)DNF (t)(GNF ,emb.(t)[kg CO2eyr]+ENF (t)[kWhyr]·EF[kg CO2ekWh])(5.6)= DEU,P,NF (t)[GBplatform ·yr]·(GINF ,emb.(t)[kg CO2eGB]+EINF (t)[kWhGB]·EF[kg CO2ekWh])(5.7)Again, the equation describes an allocation share, based on data traffic in this case, multiplied byan annualized GHG footprint, considering the entire fixed internet. Embodied emissions, GNF ,emb., areannualized and include all life cycle emissions aside from operational emissions, GNF ,op., which arisefrom use energy, ENF . Because a data transfer over a fixed network crosses both fixed access networksand core IP networks, fixed network emissions (both operational and embodied) are defined to be thesum of fixed access network emissions, and core IP network emissions; modeling details are providedin Appendix C. The total data traffic carried by fixed networks, DNF , is multiplied through; the ratio ofannualized embodied emissions to data traffic is called the embodied GHG intensity, GINF ,emb., measuredin kg CO2e/GB, while the ratio of annual use energy to data traffic is called the energy intensity, EINF ,measured in kWh/GB. Total emissions due to an end-use may thus be obtained by multiplying datatraffic generated by that end-use, in GB, by these intensities.Mobile network emissions follow a similar approach to that of fixed networks: emissions from mo-bile networks GEU,P,NM due to end-use EU on platform P may be calculated as follows, where DEU,P,NMis the annual traffic generated by the end-use that travels on mobile networks:GEU,P,NM(t) =DEU,P,NM(t)DNM(t)(GNM ,emb.(t)[kg CO2eyr]+GNM ,op.(t)[kg CO2eyr])(5.8)GEU,P,NM(t) = DEU,P,NM(t)[GBplatform ·yr]·(GINM ,emb.(t)[kg CO2eGB]+EINM(t)[kWhGB]·EF[kg CO2ekWh])(5.9)795.2.3 Datacenter emissionsDatacenter allocation is performed on the basis of network traffic generated, though this metric is aproxy; ideally datacenter impact of an end-use would be estimated by tabulating the number of data-centers, servers, or server workloads employed in servicing that end-use, but current limitations in dataavailability make this approach impractical; data-based allocation does have precedence in other studies[41]. The amount of data traffic entering datacenters due to the end-use, DEU,P,DC, is assumed to beequivalent to the amount generated by the platform (see Appendix Cfor justification). Thus:DEU,P[GBplatform ·year]= DEU,P,DC (5.10)Emissions due to datacenters GEU,P,DC are therefore calculated as follows:GEU,P,DC(t) =DEU,P,DC(t)DDC(t)(GDC,emb.(t)[kg CO2eyr]+GDC,op.(t)[kg CO2eyr])(5.11)GEU,P,DC(t) = DEU,P,DC(t)[GBplatform ·yr]·(GIDC,emb.(t)[kg CO2eGB]+EIDC(t)[kWhGB]·EF[kg CO2ekWh])(5.12)5.2.4 Overall emissionsConsidering device, fixed network, mobile network, and datacenter emissions, overall emissions aredefined as follows:GEU,P(t) = ∑D∈PIDIPGEU,D(t)+GEU,P,NF (t)+GEU,P,NM(t)+GEU,P,DC(t)[kg CO2eplatform ·yr](5.13)The model is straightforward, but requires significant empirical data inputs. In order to calculatenetwork and datacenter impacts, it is necessary to determine energy and embodied GHG intensitiesof these systems, which require estimates of total traffic, energy consumption, and embodied GHGemissions. Device impacts require estimates of embodied emissions and unit energy consumption, aswell as installed base. Allocation shares for an end-use require an estimate of the typical annual time80spent on a platform performing the end-use, as well as the typical data traffic generated by that end-useon that platform. All of these data inputs are compiled in this study from secondary sources.5.3 Model inputsRequired model inputs can be grouped into three categories, as follows: device and platform characteris-tics, including operational energy use, embodied emissions, and installed base; network and datacentercharacteristics, including operational energy, embodied emissions, and total data traffic; and end-useparameters, including data traffic generated and time spent for each platform and end-use.5.3.1 Device and platform characteristicsOperationalenergy[kWh/yr]Embodied GHG[kg CO2e/yr]Installedbase (MM)2012 2017 2012 2017 2012 2017DevicesDesktop PC 212 193 39 36 103 90Laptop PC 61 56 19 18 130 126LCD Monitor 100 109 9 8 132 120Game Console 149 190 8 7 111 115Set top box 136 136 8 7 227 235TV set 173 152 23 21 358 371DVD / Blu-ray player 25 18 2 2 238 247AV receiver 65 65 2 2 100 104Tablets 12 12 60 55 60 120Smartphones 3 3 14 13 121 207PlatformsTV platform 340 326 32 30 358 371Tablet 12 12 60 55 60 120Smartphone 3 3 14 13 121 207Desktop PC platform 308 298 47 43 103 90Laptop PC platform 87 84 22 20 130 126Average PC platform 187 173 33 30 232 217Table 5.1: Device and platform assumptions for operational energy, embodied GHG emissions,and US consumer installed baseDevice and platform parameters are shown in Table 5.1. Operational energy use, ED, is primarily ob-81tained from a Consumer Electronics Association survey [5] and laboratory measurements [217]. Deviceembodied GHG emissions per product GD,emb. and device lifespan LD, which together determine annualdevice embodied GHG emissions, are based on a review of prior LCA results developed researchers atLBNL [42]. Installed base estimates ID are based on Consumer Electronics Association estimates from2010 [5], while forecasts to 2017 for PCs, smartphones, and tablets are based on forecasts from IDC[218], Forrester [111], and eMarketer [112], respectively; for all remaining devices, installed base isassumed to scale with population, i.e. penetration rates are fixed. Calculation details including basisfor forecasts to 2017 are in Appendix C. Platforms are defined as follows. The TV platform containsone TV set and a proportion of set top boxes, game consoles, DVD/Blu-ray players, and A/V receivers;the installed base of the TV platform is equal to the number of TV sets. The Desktop PC platform isassumed to include one desktop PC and 0.97 LCD monitors, while the laptop PC platform is assumedto include one laptop PC and 0.26 LCD monitors, using results from CEA [5]. The average PC platformis weighted sum of all desktop PCs, laptop PCs, and LCD monitors; its installed base is equal to thecombined installed base of desktop and laptop PCs.5.3.2 Network and datacenter characteristicsFor each of fixed networks, mobile networks, and datacenters, the model requires energy intensity EIin kWh/GB and embodied GHG intensity GIemb. in kg CO2e/GB for each of 2012 and 2017; intensitiesare calculated by dividing total embodied GHGs Gemb. and total energy E(t) by total data traffic; trafficon fixed networks DF and mobile networks DM is obtained from the Cisco Visual Networking Index[103], a major data source for this study.The use of energy intensity as an impact factor for network data transfer was pioneered by Koomeyin 2004 [97] and is now widespread. Energy intensity is a useful metric, but it is fast-moving; Taylor andKoomey showed an order-of-magnitude reduction in internet energy intensity from 2000 to 2006 [128].It is thus of critical importance to ensure temporal consistency when comparing energy and data factors.In addition, the top-down method used by Koomey and applied here, in which total internet energy isdivided by total internet traffic, is by necessity coarse; energy intensity may also be calculated from thebottom up, by tabulating the energy per unit of traffic in smaller systems of equipment (e.g. collectionsof routers and servers), but this can lead to much smaller intensity estimates. This study’s estimates arecompared against others in Appendix C; see also a recent review [72].82llllllllllllllll l ll lllllEnergy E [TWh/yr] Traffic D [EB/yr] Embodied emissions  G_emb [Mt CO2e/yr]Energy intensity EI [kWh/GB] Embodied GHG intensity  GI_emb. [kg CO2e/GB] Overall GHG intensity GI_T [kg CO2e/GB]6.310. 2017 2012 2017 2012 2017ll Fixed networks Mobile networks DatacentersEstimated network statistics due to US consumer workloadsFigure 5.2: Network and datacenter energy, emissions, and intensities (note log scales).Total embodied energy and total energy are derived from several sources discussed in Appendix C.Figure 5.2 shows the calculated energy, emissions, and intensities for each of mobile networks, fixednetworks, and datacenters; an additional parameter, GIT , the overall GHG intensity, is obtained byadding embodied GHG intensity to the product of energy intensity and grid emissions factor.The above inputs allow for estimation of US network and datacenter intensities for 2012 and 2017.Declining trends are evident: mobile network electricity intensity drops from 3.8 kWh/GB in 2012to 0.40 kWh/GB in 2017, while fixed network electricity intensity drops from 0.12 kWh/GB in 2012to 0.060 kWh/GB in 2017, and datacenter electricity intensity similarly drops from 0.32 kWh/GB to0.15 kWh/GB. The order of magnitude gap between mobile and fixed network electricity intensities isconsistent with other findings [219] and will imply higher environmental emissions from end-uses thatuse mobile networks for data transfer, especially in 2012; next-generation mobile networks are muchless energy intense per unit of data, so the gap between mobile and fixed networks shrinks by 2017.Previous top-down estimates of network energy intensity have been higher than the results reportedhere, such as 1.8 kWh/GB in 2008 [220], or 7 kWh/GB in 2008 including end-devices [39]; this is to83be expected as energy intensity has declined over time. Bottom-up estimates tend to be smaller, sincethe system boundary is more precisely drawn and will exclude some energy is included in this study’stop-down approach; recent estimates ranging from 0.05 kWh/GB [41] to 0.2 kWh/GB [221] have beenreported, with the variation due to differences in system boundary and data year [72]. The very steepdecline in energy intensity between 2012 and 2017 shown here serves as a caution to researchers toensure that energy and emissions estimates are temporally consistent with data flows when calculatingimpacts due to network end-uses.5.3.3 End-use data traffic and time spentData traffic generated by each platform is derived from the Cisco VNI [103], shown below in Table 5.2.The platforms are disaggregated according to connection type, as there will be different network emis-sions for an end-use depending on whether the platform has a mobile network connection or not. To-tal monthly traffic for a platform is assumed to be equivalent regardless of connection type, i.e. allconnection-platform combinations generate the same amount of monthly traffic in total, though the dis-tribution across fixed and mobile networks may differ. All platforms are assumed to have a connectionto fixed networks, either via wired or Wi-Fi links. All smartphones are assumed to have access to mobilenetworks in addition to Wi-Fi. About 40% of tablets in 2012 possessed both mobile and Wi-Fi radios,while the remainder were Wi-Fi only. Finally, a small number of laptop PCs have access to mobileconnections via USB radio devices. Detailed calculations of data traffic for each platform, includingrelative installed base of the different connection types, are in Appendix C.Data traffic is further disaggregated into traffic due to online video, and due to other online uses,using data from Cisco [103], with these results shown in Table 5.3. In circumstances where platformsgenerate traffic on both fixed and mobile networks, the traffic shares are assumed to be consistent acrossall end-uses; i.e., if 60% the average smartphone’s traffic is transmitted via mobile networks, then 60%of smartphone online video traffic and 60% of smartphone other online traffic is also assumed to betransmitted via mobile networks. With these assumptions and data, it is possible to calculate data trafficon fixed and mobile networks, DEU,P,NF and DEU,P,NM , for each end-use and platform.Time spent on each end-use on each platform is also shown in Table 5.3, in hours per month. Timeestimates for online video were derived from online video traffic estimates along with assumptionsabout typical data traffic rates. Time spent online is derived from a range of market research reports,84Data traffic, 2012[GBmonth]Data traffic, 2017[GBmonth]Fixed Mobile Total Fixed Mobile TotalTV platform 16.3 - 16.3 40 - 40PC platform, mobile-connected 17.1 4.9 22 30.1 9.1 39.3PC platform, fixed/Wi-Fi only 22 - 22 39.3 - 39.3PC platform, average 21.7 0.3 22 38.4 0.9 39.3Tablet, mobile-connected 3.6 1.1 4.7 29.9 8.4 38.8Tablet, fixed/Wi-Fi only 4.7 - 4.7 38.8 - 38.8Tablet, average 4.4 0.3 4.7 35.4 2.9 38.8Smartphone 0.6 0.8 1.4 5.8 3.3 9.1Table 5.2: Estimated US monthly consumer data traffic per platform instance across fixed andmobile networks.DEU,P[GBmonth]TEU,P[Hrsmonth]Platform / end-use 2012 2017 2012 2017PC platformOffline 0 0 27.2 8.4Online video 14.9 29.3 14.5 19.4Other online 7.1 9.9 48.9 56.6SmartphoneOffline 0 0 25.8 7.7Online video 0.9 6.4 2.6 9.3Other online 0.6 2.7 36.1 60.3TabletOffline 0 0 23.9 7.2Online video 3.1 28.3 10.2 36.6Other online 1.6 9.9 25.7 28TV platformBroadcast TV 0 0 137.2 123.9IP video on demand 14.5 33.4 18 27.7Offline 0 0 3.3 3.3Online video 1.2 5 1.6 4.9Other online 0.6 1.7 1.6 1.6Table 5.3: Estimated monthly data traffic DEU,P and time spent TEU,P per platform instance, foreach end-use and platform.85notably studies from Nielsen [98], eMarketer [112], and comScore [222]; however, results from suchstudies vary widely, likely due to differences in scope and assumptions. Time estimates for all end-useswere tuned to be as consistent as possible with best available data sources; full calculation details arein Appendix C. The end-uses above encompass all active end-uses for each platform, so that for eachplatform P listed in Table 5.3, Ttotal,P is the sum of the time spent on all listed end-uses.5.4 AnalysisThe above model inputs make it possible to calculate GHG emissions due to device, network, anddatacenter for each of the study’s platforms and end-uses. The analysis proceeds by first consideringimpacts at a platform level, for all end-uses combined, using the monthly data traffic results in Table 5.2to calculate network and datacenter emissions. These results are then disaggregated into emissions dueto specific end-uses by applying time and data traffic amounts in Table 5.3. The results of the model,which give emissions estimates for each end-use on each platform at each level (device, network, anddatacenter), are aggregated and presented in several different ways in order to identify the overridingcharacteristics of US consumer ICT usage.5.4.1 Emissions by platformFigure 5.3 shows the total GHG emissions per year of using each platform in total, on a per-platformbasis and accounting for all US consumers. The graphs are colour-coded by the location to which emis-sions are attributed, which is either in the devices, in mobile networks, fixed networks, datacenters, orbroadcast TV networks. All emissions include both operational and embodied emissions, except broad-cast TV networks which exclude embodied emissions as they are expected to be negligible. Emissionsper platform are highest for televisions, desktop PCs, and mobile-connected laptops; the latter is par-ticularly high in 2012 because mobile-connected PCs generate greater amounts of mobile traffic thanmobile-connected tablets and smartphones [103], and mobile networks are much more energy intensivethan fixed networks. Forecast declines in mobile network energy and GHG intensity by 2017 reduce theenvironmental penalty of mobile-connected devices relative to fixed/Wi-Fi–only devices. Overall emis-sions among US consumers are dominated by emissions due to TV platforms, due to their relativelylarge unit emissions and very large prevalence, with most of these due to devices in consumer homes,though average device emissions are expected to decline slightly through 2017 as more energy-efficient862012 20170100200300TV platformDesktop PCLaptop (Mobile)Laptop (Wi−Fi)Tablet (Mobile)Tablet (Wi−Fi)SmartphoneTV platformDesktop PCLaptop (Mobile)Laptop (Wi−Fi)Tablet (Mobile)Tablet (Wi−Fi)Smartphonekg CO2e/yearGHG emissions per platform, US consumers2012 20170306090TV platformDesktop PCLaptop (Mobile)Laptop (Wi−Fi)Tablet (Mobile)Tablet (Wi−Fi)SmartphoneTV platformDesktop PCLaptop (Mobile)Laptop (Wi−Fi)Tablet (Mobile)Tablet (Wi−Fi)SmartphoneMt CO2e/yearTotal GHG emissions, US consumersDevices Mobile network Fixed network Datacenter Broadcast TV networkFigure 5.3: Emissions per platform and US consumer total, for all platforms, including devices,networks, and datacentersLCD televisions gain higher market share [5].5.4.2 Emissions by end-useFigure 5.4 shows a disaggregation of emissions data from Figure 5.3 according to end-use. In 2012,about 70% of emissions on the TV platform can be attributed to consumption of broadcast television; ofthese, 93% arise due to on the devices in the home. IP video on demand accounts for 25% of TV platformemissions, but this rises to 30% in 2017, with total emissions due to IP video on demand growing by30% in that time; for this end-use, device emissions account for only 30% to 40% of emissions, withthe remainder due to network and datacenters. All devices show network and datacenter emissionsexceeding device emissions for online video; all devices likewise show the opposite for other onlineactivities. Impacts per tablet platform are forecast to increase by about 50% by 2017, with the majorityof that growth due to network and datacenter emissions due to online video.Emissions across all platforms and end-uses are summed in Figure 5.5, which shows the location ofemissions for each end-use across all platforms, and in Figure 5.6, which the relative overall emissionsof each end-use, including device, networks, and datacenters, on each platform. Figure 5.5 demonstrates87TV platform PC platform Tablet Smartphone05010015020005010015020020122017Broadcast TVIP video on demandOnline videoOther onlineOfflineOnline videoOther onlineOfflineOnline videoOther onlineOfflineOnline videoOther onlineOfflineGHG emisisons, kg CO2e/yr per platformGHG emissions by end use, average per platformTV platform PC platform Tablet Smartphone0204060020406020122017Broadcast TVIP video on demandOnline videoOther onlineOfflineOnline videoOther onlineOfflineOnline videoOther onlineOfflineOnline videoOther onlineOfflineGHG emisisons, Mt CO2e/yrGHG emissions by end use, all US consumersDevices Mobile network Fixed network Datacenter Broadcast TV networkFigure 5.4: Emissions per platform, by end-usethat the majority of US consumer emissions due to digital media end-uses arise due to devices in thehome. Overall emissions grow slightly by 2017, led by an expansion in emissions due to online videoand IP video on demand, most of which arises in networks and datacenters. Emissions due to broadcastTV are expected decline slightly following a transfer of viewing time away from broadcast TV in favourof IP video on demand and online video.5.4.3 Uncertainty analysisSeveral types of uncertainty are inherent in the model. First, the model structure is dictated by sev-eral modeling assumptions, such as the allocation approach, to which the output results are sensitive.Second, each model input parameter is an estimate; while every effort has been made to obtain accu-rate estimates for each parameter and confirm their accuracy against other data sources, in some cases882012 20170204060Broadcast TVIP video on demandOnline videoOther online OfflineBroadcast TVIP video on demandOnline videoOther online OfflineMt CO2e/yearGHG emissions by end−use All US consumers, all platforms0501001502012 2017Mt CO2e/year DevicesMobile networkFixed networkDatacenterBroadcast TV networkGHG emissions, all US consumers All platforms and end−usesFigure 5.5: Emissions across all platforms, by end-use and overall, US consumer total2012 20170306090TV platformPC platform TabletSmartphoneTV platformPC platform TabletSmartphoneMt CO2e/yearGHG emissions by platform, US consumers Includes device, network, datacenter0501001502012 2017Mt CO2e/year Broadcast TVIP video on demandOnline videoOther onlineOfflineGHG emissions, all platforms, US consumersIncludes device, network, datacenterFigure 5.6: Emissions across all end-uses, by platform and overall, US consumer total89due to poor data availability the accuracy of each input parameter is difficult to assess. Third, modelparameters estimate empirical quantities to varying degrees of precision. Structural uncertainty is dis-cussed with a comparison to other studies, with suggestions for future studies to explore alternativemodels. Scenario analysis is employed to explore the sensitivity of the model to a large change in oneinput parameter. Finally, probability distributions are imposed on the model parameters to estimate theirprecision according to our subjective judgment; precision of model outputs is obtained through MonteCarlo simulation, assuming that the model structure and nominal parameter values are fixed.Structural uncertaintyStudies which make different modeling assumptions may have significantly different results. The mostclosely related study [41], estimates that consuming 10 minutes of video content at a bit rate of 2.25Gbps (corresponding to data transfer of about 170MB) would require about 3 Wh of energy due toservers plus 6 Wh if transferred via Wi-Fi or 40 Wh if transferred via 3G. According to our model,such a data transfer in the US in 2012 would incur 54 Wh due to servers plus 20 Wh due to fixed/Wi-Fidata transfer or 640 kWh due to data transfer via mobile networks, though these numbers respectivelydecline to 25 Wh, 10 Wh, and 67 Wh by 2017. Both studies agree on the relatively high impact of datatransfer via mobile networks, but the magnitude of the estimates differ considerably, largely because thatstudy’s bottom-up methodology results in smaller estimates for energy intensity than those derived here(see derivation of each input parameter in Appendix C, including comparison to [41]). Appropriatenessof modeling assumptions depends on study goals; our study is focused on characterizing aggregateUS consumer emissions and thus draws an wide boundary which may be more inclusive than those ofbottom-up studies of specific services.This study applies a US average grid emissions factor to all operational energy consumption, whichdoes not capture regional variation in emissions factors that can significantly change the emissions dueto a particular datacenter [223]. Some firms attempt to preferentially incorporate renewable energysources into their datacenter energy mix, such as Google who report their average emissions factor tobe 0.34 kg CO2e/kWh [224], about 45% less than the US average. Analysis of specific services tiedto specific datacenters should consider the local emissions factor and efficiency of those datacenters.However, given this study’s aggregate focus, the three connected end-uses – IP video on demand, onlinevideo, and other online – were assumed to be equivalent in terms of datacenter emissions intensity, all90using the US average. More detailed analysis identifying the characteristics of specific services anddatacenters which service these end-uses would reduce uncertainty regarding their emissions.In addition, this study applies a US average approach for calculating and allocating emissions whichobscures significant heterogeneity in both use behavior and deployed infrastructure. Considering dat-acenters, data traffic is used as a proxy for datacenter usage, on the assumption that traffic generatedby an end-use is correlated to energy consumed in servicing that end-use. A more direct allocation,such as number of servers and storage devices used, or number of server workloads required, would bepreferred, if data inputs were available to support such an approach, and could conceivably alter the re-sults. In particular, because video end-uses dominate total network traffic, they also dominate datacenteremissions under a data allocation model, but it is not known if the majority of datacenter emissions areactually dedicated towards servicing video. If the relationship between data traffic and datacenter energyis not linear, then this study could overestimate the datacenter component of emissions due to video, andunderestimate impacts due to other online services. It is also possible that some ICT applications, whichmay or may not fall within the categories assessed in this study, require highly disproportionate data-center resources relative to network traffic generation. Further research is required to better understandthese relationships. In the mean time, a simple sensitivity analysis is performed below.Parameter accuracyDatacenter energy intensity was difficult to assess due to lack of data as discussed in Appendix C. Asector-wide intensity of 0.32 kWh/GB in 2012 and 0.15 kWh/GB in 2017 was applied in this study, butGoogle’s self-reported datacenter energy in 2012 was 0.08 kWh/GB [224]. While the average datacen-ter will be less efficient than Google’s, which are among the industry’s most efficient, it is possible thisstudy overestimates average datacenter energy of consumer ICT services. Accordingly, an “efficient dat-acenter” scenario was created which applies Google’s datacenter energy intensity uniformly, assumingit declines to 0.04 kWh/GB in 2017. The implications of this analysis, shown in Figure 5.7, are straight-forward; a 75% reduction in datacenter energy intensity results in a corresponding 75% reduction indatacenter emissions.Because the model is linear and multiplicative, sensitivity analysis is straightforward; for example, a10% change in network energy intensity would result in a 10% change in network energy consumptionof an end-use. The model specification and input parameters provided in Sections 5.2 and 5.3 fully912012 20170100200300400TV platformDesktop PCLaptop (Mobile)Laptop (Wi−Fi)Tablet (Mobile)Tablet (Wi−Fi)SmartphoneTV platformDesktop PCLaptop (Mobile)Laptop (Wi−Fi)Tablet (Mobile)Tablet (Wi−Fi)Smartphonekg CO2e/yearGHG emissions per platform, US consumers Baseline scenario2012 20170100200300400TV platformDesktop PCLaptop (Mobile)Laptop (Wi−Fi)Tablet (Mobile)Tablet (Wi−Fi)SmartphoneTV platformDesktop PCLaptop (Mobile)Laptop (Wi−Fi)Tablet (Mobile)Tablet (Wi−Fi)Smartphonekg CO2e/yearGHG emissions per platform, US consumers Efficient datacenter scenarioDevices Mobile network Fixed network Datacenter Broadcast TV networkFigure 5.7: Emissions per platform and US consumer total, baseline and efficient datacenter sce-narios.describe the model, such that it may adapted for use in other contexts in which the values of some inputparameters may change.Parameter precisionModel precision is assessed assuming that all modeling assumptions are fixed; i.e. all of the assumptionswhich were used in calculating model input parameters are assumed to be appropriate. Approximateconfidence intervals are applied on each empirical input parameter to express the range within the truevalue lies, in our best subjective judgment, taking into account the expected precision of the underlyingdata sources. Many parameters describe populations with high variability or heterogeneity, such asannual energy consumption of a desktop PC among US consumers; because the study is concerned withaggregate impacts, the parameters are estimates of the population mean. As such the confidence intervalsdo not capture the significant variability within the population. Consequently, the average emissionsper device reported in this study apply to a fictitious average consumer and may be unrepresentative formany categories of consumer with behavior far from the mean; future work could explore how variabilityamong consumer usage patterns affects emissions, perhaps stratifying the population into light, medium,92and heavy users of a particular platform or service. Such a study would require behavioral data at a finerlevel of disaggregation than the sources applied here. Table 5.4 shows the 95% confidence intervalswhich were applied to each input parameter; justifications are in Appendix C. To illustrate the impactsof uncertainty on model outputs, a 1000-trial Monte Carlo simulation was performed, assuming normaldistributions on each model input, with resulting error bars shown in Figure 5.8, where the error barsspan ± two standard deviations from the nominal output.95% C.I.2012 2017Network and datacenter energy intensities, EI ±20% ±40%Network and datacenter embodied GHG intensity, GIemb. ±30% ±50%Device energy ED ±20% ±30%Device embodied emissions GD,emb. ±20% ±30%Device installed base ID ±5% ±10%Time share per end-use on each platform TEU,P ±15% ±30%Traffic per end-use on each device DEU,P ±10% ±20%Table 5.4: Estimated confidence intervals on model input parameters.It is possible to conclude that emissions due to devices significantly exceed those due to networksand datacenters, but the distinction between datacenter, fixed network, and mobile network emissions isless strong, especially in 2017, and particularly if one accounts for the possible differences in modelingassumptions such as the efficient datacenter scenario discussed above. Likewise, it possible to con-clude that emissions due to broadcast TV are largest, and that emissions due to the three video end-uses(broadcast TV, IP video on demand, and online video) significantly exceed remaining emissions, but thedifferences between IP video-on-demand, online video, and other online end-uses are not significant.By 2017, the lower-range of the confidence interval for broadcast TV emissions overlaps with the upperrange of emissions estimates for other end-uses, reflecting uncertainty in usage patterns looking for-ward, though the most likely outcome is that broadcast TV emissions will remain the largest. Improvedprecision would require higher-quality data sources regarding user behavior with respect to time spentand data traffic generated due to each end-use. Finer-grained disaggregation into a larger number ofend-uses would likely be challenged by additional uncertainty caused by the incorporation of new datasources.932012 20170255075Mt CO2e/yearBroadcast TVIP video on demandOnline videoOther onlineOfflineGHG emissions, total US consumers By end−use2012 2017050100150Mt CO2e/yearDevicesMobile networkFixed networkDatacenterBroadcast TV networkGHG emissions, total US consumers By location of emissionsFigure 5.8: Monte Carlo results showing uncertainty range for GHG emissions, US consumer to-tal; error bars span ± two standard deviations.5.4.4 Energy comparison to household appliancesIn Figure 5.9, the operational energy only in kWh/year of using each platform is compared to the annualconsumption of various electric household appliances according to DoE data [225]; network and data-center energy use are grouped together. A recent controversial report published online by Mills claimedthat network and datacenter energy due to an iPhone could exceed that of a refrigerator [226], thoughthis was refuted online by Koomey [227]. Our model suggests that a smartphone operated in the USconsumes at most about 36 kWh/year in 2012 and about 39 kWh/year in 2017, including network anddatacenter energy, which is about 5% of a typical refrigerator’s annual usage of 620 kWh/year. Theannual energy use of a tablet is higher, at about 37 kWh/year (Wi-Fi) and 110 kWh/year (mobile) in2012, growing to 84 kWh/year (Wi-Fi) and 143 kWh/year (mobile) in 2017, the vast majority of whichoccurs in networks and datacenters. These levels of combined device and network energy consumptionplace tablets in a similar category as some appliances like microwaves and dishwashers, but still farbelow desktop PCs and TV platforms.942012 2017SmartphoneTablet (Wi−Fi)Hair dryerToaster ovenCoffee makerTablet (Mobile)Ceiling fanClothes washerDishwasherMicrowaveLaptop (Wi−Fi)Laptop (Mobile)Desktop PCTV platformFridge/freezerPool pumpClothes dryerWater heater0 1000 2000 0 1000 2000Typical annual energy use (kWh/year)Networks and datacenters DevicesFigure 5.9: Comparison of ICT energy use against other electric household appliances5.5 ImplicationsLife cycle GHG emissions due to digital media end-uses across different platforms were calculatedusing estimates of network and datacenter energy and embodied GHG intensities, which were applied toa model of data traffic and time spent on each end-use and platform for 2012 and 2017 in a US consumercontext. Among end-uses, broadcast television is the largest contributor to emissions, accounting for42% of total US consumer emissions due to ICT end-uses in 2012 and 34% in 2017, the vast majorityof which occur due to devices in the home. IP video-on-demand, another end-use which is consumedon the TV platform exclusively, represents a further 16% in 2012 and 19% in 2017; in contrast tobroadcast television, the majority of emissions due to IP video-on-demand occur due to network anddatacenter infrastructure, largely because this delivery mechanism requires a high amount of traffic toindividual subscribers, unlike broadcast television which spreads the infrastructure burden over manymore viewers, confirming results in a BBC study [45].Viewing of video content on a TV platform is currently the most popular consumer ICT end-use interms of hours spent, and this seems likely to continue; due to the relatively high energy use of TV setsand associated peripherals and their popularity, along with the relatively high data traffic rates required todeliver video, video on TV platforms is expected to remain the largest contributor to emissions through2017. The potentially disruptive traffic burden due to an ongoing transition from broadcast to on-demand95has already been noted [100], but this transition will have an impact on emissions as well, in effectincreasing the network and datacenter burden of consuming TV content. The effect, though modestunder this study’s assumptions, is proportional to the degree of substitution of video on-demand forbroadcast TV, which could grow substantially over time. The best opportunities for reducing impactsare thus reducing the impacts of the TV platform, especially the energy consumed by televisions,set-top-boxes, and game consoles, reducing the network load required to deliver on-demand video content.Network operators already have strong market incentives to meet to latter objective; the former canbe targeted through policy initiatives such as ENERGY STAR. In addition, the next-highest portionof overall emissions after TV platforms is due to desktop PC devices. According to a 2010 plug-loadaudit, two-thirds of desktop PCs are left on overnight or for large periods of time, despite the availabilityof sleep or hibernation modes which were often turned off and which could cut aggregate desktop PCenergy use in half if widely used [228]. These findings are not new; energy researchers have previouslyidentified the prominent energy consumption of TV sets, set-top boxes, and desktop PCs in the home[229] and the energy savings opportunities of power management on these devices [228]. Our studysuggests that this energy consumption remains of primary importance even when network and datacenteremissions are taken into account.Despite this finding, there is value in accounting for network and datacenter emissions, especially fordevices with lower operational energy consumption. A life cycle assessment of a tablet device that doesnot consider network or datacenter impact would account for only 30% to 50% of the emissions thatarise due to end-uses performed on the device. However, reports that an iPhone’s energy use includingnetworks and datacenters could be comparable to that of a refrigerator [226] are not supported by ourstudy. Emissions due to mobile networks from mobile-connected devices were particularly high in2012, especially for mobile-connected laptops which generated large amounts of mobile traffic, but theongoing shift toward next-generation networks will greatly improve mobile network emissions intensityallowing emissions due to mobile networks to decline.Emissions due to network-enabled ICT end-uses are a moving target. While network and datacenterefficiencies do improve dramatically over time, total traffic demand also grows at a comparable rate,leading to a modest net gain in emissions due to networks and datacenters. The emissions due to someend-uses may decrease over time on a functional unit basis, such as the emissions due to downloadingan hour’s worth of music, which would decline over time so long as the data rate of the music stays96constant. However, expected 280% growth in total US consumer data traffic from 2012 to 2017 [103]indicates some combination of increasing consumption of end-uses, increasing data rates of end-uses(e.g. higher quality video), and/or emergence of new end-uses (e.g. ambient video streams) duringthat time. It is difficult to draw any general conclusions about the broader impacts of ICT withoutconsidering these trends. Future research on this topic will depend on the continuing availability offrequently refreshed data sources describing device installed base, energy consumption and emissions;network and datacenter energy and emissions intensities; and user behavior with respect to end-uses.Further disaggregation into more detailed end-uses such as social networking, email, e-book readingetc. would be possible with better data sources indicating time spent and data traffic generated by theseend-uses; market research reports can be a valuable data source for understanding user behavior, butare often not freely available, and may not fully report study assumptions leading to unexplainabledisagreements with other studies.Overall, video is the dominant category of ICT end-use among US consumers, accounting for aboutthree-quarters of total emissions, including broadcast television, IP video on demand, and online videoacross all platforms. Because of its high data rates, video traffic represents about 80% of total IP trafficaccording to Cisco [103], and thus accounts for about 80% of network and datacenter emissions undera data-based allocation model; it accounts for a large share of emissions due to electronics in the homeas well primarily due to the large impacts of TV sets and associated peripherals. Thus, understandingconsumer behavior with regards to video is of critical importance toward understanding emissions dueto consumer use of ICT.97Chapter 6ConclusionThis chapter contains a summary of specific findings from each of the previous three chapters; a discus-sion of the limitations of the research; suggested future work, both in terms of methodological improve-ment and study topic; recommendations for practical application of this work; and commentary aboutbroader issues in the domain of ICT and environment.6.1 Specific findings and significanceSpecific conclusions from each of the studies in Chs. 3, 4, and 5 are summarized below. In addition, thesignificance of each study is discussed in relation to existing published research in this domain.6.1.1 Life cycle assessments of desktop PCsThe meta-analysis of previous LCA studies of desktop PCs resolved some previous ambiguities in thisliterature, most notable a disagreement regarding which phase of the product life cycle was dominantin terms of energy and GHG emissions – production, or operation. Through a thorough review of priorstudies of operational energy consumption of desktop PCs, including field measurements, surveys, anddata from ENERGY STAR, the study established that a reasonable range for operational energy con-sumption of a desktop PC was between 100 and 350 kWh/yr, with overall mean energy consumptionlikely towards the middle of that range. Two of the studies which showed higher energy and emissionsdue to production rather than operation made unrealistically low assumptions of operational energyconsumption, namely, 54 kWh/yr for Williams [13] and 49 kWh/yr for Choi et al [167]; Williams alsoestimated PC lifespan to be 3 years, which is on the lower end of the range used by other studies,98and accounted for only use electricity rather than primary energy due to electricity. Taking these dis-agreements into account, the study concludes that under typical, average usage conditions, energy andemissions due to operation of a desktop PC usually exceeds those due to its production, and that thestudies which reached the opposite conclusion did so after making unreasonably low assumptions re-garding operational impacts. In addition, the study showed that differences in product from factor, e.g.between a large workstation-class desktop and a small integrated desktop, could be a dominant factorinfluencing the study result.The study highlights the difficulty in assessing and comparing published LCA studies which do notprovide full details of modeling assumptions and source data, as only a small number of studies couldbe decomposed in sufficient detail to allow for comparisons of modeling assumptions; others werecompletely non-reproducible, which is troubling. The comparison showed that studies were sensitiveto assumptions regarding operational behavior among end-users, as well as to differences in productinventory and source data, much of which is not provided in many studies. As such it is difficult toknow that the results of any LCA study in particular are valid, or the circumstances under which theymay be valid, aside through processes of extensive meta-analysis as I demonstrate in this study.6.1.2 Embodied emissions of electronicsStudy of the embodied emissions of 11 electronic device specimens yielded three primary conclusions.First, modern desktop, laptop, and LCD monitor specimens had embodied GHG emissions about 50%to 60% lower than those of earlier-generation products originally modeled in ecoinvent, which canprimarily be attributed to reduced usage by mass of circuit boards and integrated circuits; this is aconsequence of Moore’s Law, which has allowed for higher levels of functionality to be achieved onfewer integrated circuits. Second, embodied emissions showed a roughly linear trend with respect tomass: to first order, heavier products had higher levels of embodied emissions, a trend which was alsoobserved in Apple’s product environmental reports [212]. Third, a statistical model was developedwhich enabled first-order emissions estimation using only product mass, i.e. 27 kg CO2e/kg if themodel is fit using my study’s data, or 37 kg CO2e/kg if the model is fit using Apple’s data; a moresophisticated model using the masses of the display, circuit board, and battery was also developed andproduces better results, especially for lower-mass products.The study is unique in its scope; aside from the Apple product environmental reports, which are not99peer-reviewed, no other study has included a comparison of this number of products. The integratedapproach of modeling all products within a consistent framework allows us to draw general conclusionsabout the characteristics of electronics products, i.e. that embodied emissions are lower for compact,lightweight models, which can be used as heuristics for decision-making. The statistical modelingapproach towards impact estimation, though in need of further validation and development, may offer ameans of characterizing environmental impacts of a wide range of products without need for individualdetailed LCA studies of each.6.1.3 Emissions of connected mediaA model of GHG emissions due to end-uses in the home showed that video was the dominant end-use from an emissions perspective, both in the home due to high energy use from TV sets and relateddevices, and in networks and datacenters due to high amounts of data traffic attributable to video. Thenetwork and datacenter components of emissions (collectively, the ‘network emissions’) were smallerthan device emissions overall. On mobile devices, network emissions far exceed device emissions,especially when consuming video content, and thus significantly enlarge the GHG footprint of using thedevices. However, even considering network emissions, the overall emissions due to mobile products aredwarfed by those of larger devices like desktop PCs and televisions, for which the majority of emissionsarise due to devices. Conjecture that an iPhone’s network emissions could exceed the emissions ofoperating a refrigerator was not supported. Aggregate emissions due to consumer ICT use in the USis expected to remain relatively stable into 2017, though total emissions due to devices are forecast todecline slightly (by about 10%) while network emissions grow slightly during that time.The integrated, top-level approach of this study is unique, as is its breadth; no other study hascompared the impacts of broadcast TV, video-on-demand, and other online and offline end-uses, andno other study has compared the impacts of end-uses using TVs against those using PCs, tablets, andsmartphones. The most closely related work, by Schien et al [41, 132] examined a subset of these end-uses and devices, and estimated aggregate impacts of one service only, namely online news delivery.The integration of behavioral data from market research firms was a key innovation in the study inChapter 5 enabling aggregate estimates of multiple end-uses. The study helps us understand the impactsof connected media consumption in many different forms, and provides a template for characterizingthe impacts of cloud services which are growing to dominate the ICT landscape.100These conclusions should be considered alongside the methodological limitations of the study,which are outlined in the following section.6.2 LimitationsThe research in this dissertation is constrained by some general limitations, discussed below. In addition,each of the three studies has unique issues related to methodology and approach, which are outlined inturn. These limitations motivate a discussion of future research needs which follows in the next section.6.2.1 General limitationsThe studies are constrained by the limitations of LCA in general, which were discussed in Section 2.4.The studies rely on average data and models, especially for upstream industrial processes; and they aresensitive to modeling assumptions, input data source, and aggregation methodology. The tension be-tween results obtained from process-sum LCA, hybrid LCA, and EIOLCA was not directly confrontedin this research program. Because of this limitation, the research in this dissertation focus on compar-isons within a framework, rather than the reporting of absolute numbers.The studies consider impacts only in terms of GHG emissions and (in some cases) primary energydemand, and focus primarily on operational and manufacturing phases which dominate in those impactdimensions. While these are operationally useful dimensions of impact, they do not capture many otherimportant forms of environmental impact which are not considered in this dissertation. Thus, when thestudy results discuss lower-impact products or behaviors, this must be understood to refer only to thoseimpact dimensions specified.In addition, the studies all considered a US context exclusively, with a US grid mix. Other ge-ographic jurisdictions would naturally have different grid emissions factors, which would affect theoperational impacts. Because ICT is a highly commodified industry, production emissions of ICT de-vices identified through LCA can be considered to be a global average, and thus not affected by studygeographic scope. However, behavioral patterns in non-US contexts may differ considerably, as couldpenetration rates for different devices, especially in less-developed countries; network energy and emis-sions intensity may differ as well.1016.2.2 LCA meta-analysisThe study in Chapter 3 attempted to draw conclusions about the GHG emissions of a typical desktopPC under typical operating conditions in the US, but this is a simplification which obscures the highamounts of variation possible both in product form and function, and patterns of usage; the study’sconclusion, that operational emissions usually exceed emissions due to manufacturing, will not hold forall types of desktop or pattern of usage. In addition, the conclusions regarding plausible ranges for adesktop PC’s GHG emissions due to manufacturing are tentative, based on the very small pool of studiesfrom which they were derived.6.2.3 Product LCA and embodied emissionsTruncation error is a particular issue in the study in Ch. 4, which is a likely contributing reason as towhy it reports significantly lower GHG emissions than those reported by Apple for comparable prod-ucts. Some substances were not easy identifiable using the teardown method employed in Ch. 4, suchas coatings, metallic alloys, or blended plastics, and would have thus been excluded. Since Apple’sanalysts had access to proprietary industrial data including full bills of materials of each products, Iwould consider their numerical results to be more accurate than those reported in Ch. 4. However,trends between products were consistent in both datasets, as was the linear relationship between productcharacteristics and embodied emissions; only the numerical values of the coefficients differed. Furthertruncation error due to upstream economic sectors not accounted for in Apple’s dataset may exist, sincethose results were produced using a process-sum methodology according to their analyst 1; there is notenough information to analyze or bound this error, since the study details are not public. I considerApple’s reports to be the best currently available data source regarding the impacts of their products, butcaution must still be exercised when applying these results outside their original context.The study in Chapter 4 considered only embodied emissions, rather than the full product life cyclewhich would include operational emissions. The study’s conclusion that lighter, smaller products havefewer embodied GHG emissions than heavier products, should hold in terms of operational emissions aswell, since mobile products tend to be very power-efficient, but the study did not assess this explicitly.1Personal communication, 20131026.2.4 Connected mediaThe study in Ch. 5 applied US consumer averages in all consideration of behavior and impacts, pri-marily due to limited data availability. This approach is well-suited towards consideration of aggregatebehavior, which was the study’s focus, but it does not capture the wide variability of many of the studymodel’s inputs, which may have highly skewed and/or multimodal distributions. Thus, some of thestudy’s primary conclusions – e.g., that most emissions are due to TV sets and related equipment – maynot hold for some categories of individuals, such as very heavy consumers of internet video on PCs.The study could be extended to include more fine-grained groupings of consumers, with the availabilityof higher-resolution data sources.Likewise, the study in Ch. 5 grouped all internet services together and considered internet-wideaverages of energy and GHG emissions. Again, this framing is useful for considering aggregate impacts,but presents a simplified picture. Data center allocation was performed according to network data traffic,which has precedence in other studies but is a simplification; ideally, the number of datacenter workloadsattributable to each class of service would be known, which would enable more accurate allocation.There are significant structural uncertainties in the allocation methods which suggest that the estimatesof end-use emissions should be taken as first-order estimates. The estimates of platform emissionswhich are not disaggregated by end-use are more certain.6.3 Future research needsThis dissertation suggests several important areas of future research, which are grouped into three cate-gories: methodological issues relating to the application of LCA to ICT products and services; knowl-edge gaps relating to ICT products; and knowledge gaps relating to ICT services and infrastructure.6.3.1 MethodsThe influence of truncation error on LCA results and the use of hybrid methods continues to be an im-portant research frontier. The surprisingly high magnitude of the economic correction obtained throughhybrid LCA in this domain is problematic, as it means that results obtained from process-sum LCA,which includes the vast majority of published work on ICT, may significantly underestimate emissions.Better data is needed regarding upstream process models for manufacturing and end-of-life, and furtherexploration of truncation error and hybrid LCA, in order to improve our understanding of device impacts103through LCA.The continued use of process-sum LCA in spite of its potential weakness may be attributed in partto its relative simplicity. As many LCAs come from industrial practitioners who tend to be resource-constrained, there is a need for streamlined, simplified tools; an increase in complexity towards hybridLCA may improve accuracy while simultaneously reducing the accessibility of the method. Thus, con-tinuing fundamental research into hybrid LCA should aim to identify simplified tools and rules-of-thumbwhich are accessible to the practitioner community.Significant differences were identified between top-down and bottom-up estimates of the GHG emis-sions of ICT end-uses, especially considering emissions due to the internet. The top-down energy andemissions intensities derived in Ch. 5 used an inclusive approach based on best available data sourcesto estimate overall energy and emissions and data traffic. The appropriate methodological choice mightdepend on the study goal; a study attempting to estimate aggregate GHGs as in Ch. 5 should use a top-down approach in order to avoid truncation error, while a bottom-up method might be more appropriatefor detailed study of specific end-uses or systems in order to identify emissions hotspots. Still, thereremains an unresolved question as to the source of the disagreement between the two methods whichshould be explored.6.3.2 ICT productsNew products continue to emerge on the market, including smart glasses, smart watches, and connectedsensors in the home; existing products continue to evolve through new technological advances suchas ultra-high-resolution displays which appear to bring a significant emissions premium according toApple’s latest product environmental reports [212]. Continuing primary research regarding the compo-sition, upstream manufacturing processes, and overall emissions of these products will be needed.When considering impacts of devices, point estimates are less reliable than integrated analysesor meta-analyses. In particular, for estimates of operational energy consumption of ICT devices, re-searchers should make use of empirical data which exists from large-scale studies of user behavior, asdiscussed in Ch. 3. Likewise, given the large number of devices currently on the market, includingvariations in form factor within a given product category, and hybrid devices which span formerly dis-tinct product categories, point estimates of the impacts of one device are of declining utility. The studyin Ch. 4 argues for a shift towards broader frameworks considering the impacts of multiple products104simultaneously. In addition, the statistical analysis in Ch. 4, though simple in its construction, sug-gests the possibility of specifying device impacts in terms of product characteristics using simple linearequations, which could expand the breadth of product LCA studies.The GHG emissions due to networks and datacenters arising from connected end-uses on ICT de-vices can be significantly larger than those due to the devices themselves in the case of small mobiledevices like smartphones, tablets, and to some extend, laptop PCs. Since normal operation of these de-vices includes extensive use of connected services, a reasonable assessment of the impacts arising fromthe operation of these devices should include the impacts due to networks and datacenters, likely usingan approach similar to that developed in Ch. ICT services and infrastructureAssessment of impacts of ICT services requires high-quality models of networks and datacenters, in-cluding information regarding their operational energy consumption and embodied emissions. The en-ergy and emissions intensities developed in Ch. 5 are coarse and rely on simplified infrastructure models,and would be improved by more focused study, especially empirical assessments of energy consump-tion. In addition, the relationship between ICT services and datacenter resources is poorly understoodand modeled only in aggregate in Ch. 5; new data sources offering a more detailed mapping of servicesto resources would enable studies of specific ICT services (e.g. Netflix, Facebook) rather than aggregatecategories. Collecting such data will be hampered by their proprietary nature. However, sustainabilityreporting from firms such as Google and Facebook has shed at least some light on industry practicethat were hitherto opaque, as has some pioneering research by Greenpeace [223]; continuing researchand data disclosure could enable more accurate and higher-resolution estimation of the impacts of ICTservices, which would support environmentally conscious decision-making by consumers of those ser-vices.The market research data used in Ch. 5 was an invaluable data source, but these studies report onlylimited modeling details which makes it challenging to resolve disagreements across studies. It seemsunlikely that peer-reviewed academic studies will be able to research emerging products and behaviorsat a speed comparable to market research firms; thus, future research of ICT services will likely dependon market research. Accordingly, research is needed to validate and calibrate these data sources, perhapsthrough academic-industry partnerships.105Because of rapid change in user behavior as well as technological improvements in network and dat-acenter efficiency, point estimates of the emissions of a specific end-use are of declining utility, echoinga finding above relating to device emissions; many of the studies reviewed in Ch. 2 analyze end-useswhich are already in rapid decline. In order to produce meaningful research with some longevity, re-searchers should move past snapshot views and consider longer-term trends. An interesting observationis that while the collection of end-uses for which devices are used changes very rapidly, device annualenergy consumption as reviewed in Ch. 5 changes relatively slowly. By tracking device installed baseand total device energy consumption instead of specific end-uses, studies may achieve higher relevancefor a longer period of time. This will come at the cost of some precision, but I would argue there is lim-ited value in achieving high levels of precision when analyzing such a fast-moving industry. High-levelintegrated frameworks are a useful approach which should be considered in future studies.6.4 ImplicationsThe research was designed to support green decision-making among ICT-using firms and organizations,as well as to inform policy-makers who are attempting to reduce emissions due to ICT. Specific recom-mendations to these actors are outlined below. In the following section, commentary is offered regardingthe ICT/environment research field in general.6.4.1 RecommendationsFor ICT-using firms, office workers should be given compact desktops, thin client devices, laptops, ortablet PCs where possible, rather than larger desktop PCs, as this results in a significant decrease inembodied GHG emissions; operational energy consumption of these devices in an office context wasnot explicitly studied in this dissertation, but compact and mobile devices tend to use significantly lessoperational energy as well. Power-savings features should be activated for all devices, but especiallydesktop PCs and large displays, which in a consumer context are dominant energy consumers; thesedevices should be turned off or in a low-power state when not in use. ICT services in an office settingwere not studied in this work, but a different study has shown that offices should prefer cloud servicesrather than custom-hosted software running on dedicated servers [42], which leads to reduced networkand datacenter emissions of running office productivity software.For ICT-using individuals, media consumed on mobile devices will lead to lower emissions than if106consumed via televisions or desktop PCs; the use of Wi-Fi connections will lead to less emissions thatthe use of mobile connections. Individuals should ensure power management features are in use on alllarge energy-using devices, including television sets, game consoles, set-top boxes, and desktop PCs,and especially that these devices are turned off or in a low-power sleep mode when not in use. Wherepossible, individuals should preferentially purchase energy-efficient devices.For labelling authorities such as the Green Electronics Council (responsible for EPEAT) and theUS EPA (responsible for ENERGY STAR), when calculating product impacts, consider incorporatingan estimate of the network and datacenter impacts attributable to the product under typical operatingconditions. Once the research has advanced sufficiently, labelling authorities should consider adoptingempirically-derived statistical models of product GHG emissions in order to quantify emissions of awide range of products.For regulatory agencies addressing the broader impacts of ICT usage, consider enforcing efficiencystandards on heavy energy-using products, similar to the California Energy Commission’s regulationsspecifying maximum power consumption ranges for televisions sold in that state [230]. Televisions,set-top boxes, game consoles, and desktop PCs are of particular importance, even when network anddatacenter energy consumption is accounted for. Procurement directives for government agencies shouldpreferentially support laptops, lightweight desktops, thin clients, and tablets over desktop PCs.6.4.2 CommentaryThough not specifically discussed within the scope of the research in this dissertation, I offer the fol-lowing commentary and conjecture as to other important areas of research, considering the question ofthe environmental impacts of ICT more broadly. First, the findings regarding desktop PC and TV equip-ment being the highest contributor of energy consumption in the home are familiar, and while these arebeing addressed though policies such as ENERGY STAR and efficiency regulations, there are limits tothe ability of such policies to drive behavior change. There is a rich literature regarding the mechanismsof behavior change with respect to energy consumption which was beyond the scope of the researchin this dissertation (e.g. [231]). Integrating such research alongside the study of energy and emissionsdue ICT end-uses is a clear next step in order to aid in the design of effective demand-side managementprograms.Second, though not discussed in this thesis, water usage is an emerging area of potentially high107importance, both in upstream manufacturing facilities, and operationally in datacenters, especially forcooling. Progress in water usage assessment has been primarily industry-led, through some water foot-print analysis by Intel [232], and through the development of the Water Usage Effectiveness (WUE)statistic for datacenters pioneered by The Green Grid, an industry consortium [233]. Likewise, im-pact dimensions other than GHG emissions and primary energy use are underdeveloped, both in theconsideration of impacts of ICT, and in the LCA field in general. Many of these dimensions may benumerically problematic as discussed in Section 2.4, thus there is a need for improved LCA models andcharacterization schemes to quantify impacts; however, regardless of the strength of numerical impactmodels, the risks of toxic releases and exposures in product manufacturing and disposal will need to bemanaged through regulatory responses, for which LCA-quantified impacts are only one of many impor-tant dimensions. The use of nanomaterials and other new chemicals is an area of particular concern; thework of Beaudrie et al is relevant here [234].Finally, the application of ICT in other sectors such as energy and transportation was identified as apotential emissions wedge by several studies [30, 235]. More accurate estimates of the emissions savingspotentials could be obtained through the use of techno-economic models to predict the ripple effects ofchanges to these industries. However, the research priority should be on identifying mechanisms toachieve the highest possible emissions savings, which are likely domain-specific; for example, higheruptake of smart grids requires policy instruments specific to the energy sector, while higher penetrationof electric vehicles requires policy instruments specific to the automotive sector. I see little commonalityamong these various applications of ICT which would suggest the possibility of ICT-centric instrumentsapplicable to each.6.5 ConclusionThe research in this dissertation improves our understanding of the energy and GHG emissions due toICT devices and end-uses through three integrative studies, which explored the life cycle GHG emissionsand primary energy demand of desktop PCs, the embodied GHG emissions of 11 electronics devices,and the GHG emissions due to five categories of end-use across four devices including emissions dueto device, network, and datacenter. The studies make a significant contribution of primary data, throughthe provision of source data from the study of embodied emissions in Ch. 4, and of methodologicalimprovements, via the statistical model of GHG emissions relating to product characteristics derived in108Ch. 4, and via the approach for incorporating behavioral data from market research in Ch. 5. In addition,the model in Ch. 5 includes a significant compilation of secondary data sources, including forecaststo 2017, which are transparently reported so as to make them easily adaptable by other researchers.The study findings, detailed above, improve the ability of ICT-using organizations and individuals topreferentially support lower-impact products and behaviors, and support policy-makers who wish toencourage such behavior.109Bibliography[1] P. Teehan and M. Kandlikar, “Sources of variation in life cycle assessments of desktopcomputers,” Journal of Industrial Ecology, vol. 16, pp. S182–S194, 2012. → pages iv, 7, 37, 59,62, 74, 141[2] ——, “Comparing embodied greenhouse gas emissions of modern computing and electronicsproducts,” Environmental Science & Technology, vol. 47, no. 9, pp. 3997–4003, May 2013. →pages iv, 7, 62, 74, 169[3] P. Teehan, M. Kandlikar, and H. Dowlatabadi, “Estimating the changing environmental impactsof ICT-based tasks: a top-down approach,” in Proc. IEEE International Symposium onSustainable Systems and Technology, May 2010. → pages v[4] K. W. Roth and K. McKenney, “Energy consumption by consumer electronics in U.S.residences,” TIAX LLC, Cambridge, MA, Jan. 2007, TIAX D5525; Final report to theConsumer Electronics Association. → pages x, 16, 17, 51, 155[5] B. Urban, V. Tiefenbeck, and K. Roth, “Energy consumption of consumer electronics in U.S.homes in 2010,” Fraunhofer Center for Sustainable Energy Systems, Cambridge, MA, 2011.[Online]. Available: http://cse.fraunhofer.org/publications/energy-consumption-of-consumer-electronics-us-households-2010/ → pages x, 1, 16, 17, 20,28, 73, 77, 82, 87, 155, 157, 161, 169[6] N. Negroponte, Being digital. Random House of Canada, 1995. → pages 1[7] D. Coyle, The Weightless World: Strategies for Managing the Digital Economy. MIT Press,Jul. 1999. → pages 1[8] D. Quah, “The weightless economy in economic development,” World Institute forDevelopment Economics Research, Report, 1999. → pages 1[9] L. M. Hilty and T. F. Ruddy, “Sustainable development and ICT interpreted in a natural sciencecontext,” Information, Communication & Society, vol. 13, no. 1, p. 7, 2010. → pages 1[10] E. Heiskanen, M. Jalas, and A. Ka¨rn, “The dematerialization potential of services and IT: futuresstudies methods perspectives,” in The Quest for the Futures seminar, Turku, Finland, 2000. →pages 1[11] B. Allenby, “E-environment: Hype and promise,” Environmental Quality Management, vol. 9,no. 4, pp. 37–43, 2000. → pages 1[12] E. D. Williams, R. U. Ayres, and M. Heller, “The 1.7 kilogram microchip: Energy and materialuse in the production of semiconductor devices,” Environmental Science & Technology, vol. 36,no. 24, pp. 5504–5510, Dec. 2002. → pages 1, 135110[13] E. Williams, “Energy intensity of computer manufacturing: Hybrid assessment combiningprocess and economic input-output methods,” Environmental Science & Technology, vol. 38,no. 22, pp. 6166–6174, Nov. 2004. → pages 1, 3, 20, 34, 43, 45, 47, 50, 51, 52, 53, 54, 55, 61,98, 134, 135, 136[14] J. Puckett, S. Westervelt, R. Gutierrez, and Y. Takamiya, “The digital dump: Exporting re-useand abuse to Africa,” Basel Action Network, 2005. [Online]. Available:http://ban.org/library/TheDigitalDump.pdf → pages 1[15] J. Puckett, L. Byster, S. Westervelt, R. Gutierrez, S. Davis, A. Hussain, and M. Dutta,“Exporting harm: The high-tech trashing of Asia,” Basel Action Network and Silicon ValleyToxics Coalition, 2003. [Online]. Available:http://www.ban.org/E-waste/technotrashfinalcomp.pdf → pages 1[16] E. Grossman, High tech trash: digital devices, hidden toxics, and human health. Island Press,May 2006. → pages 1[17] P. Huber and M. P. Mills, “Dig more coal - the PCs are coming,” Forbes Magazine, vol. 31,1999. → pages 1[18] J. G. Koomey, C. Calwell, S. Laitner, J. Thornton, R. E. Brown, J. H. Eto, C. Webber, andC. Cullicott, “SORRY, WRONG NUMBER: the use and misuse of numerical facts in analysisand media reporting of energy issues,” Annual Review of Energy and the Environment, vol. 27,no. 1, pp. 119–158, Nov. 2002. → pages 1[19] A. Andrae and O. Andersen, “Life cycle assessments of consumer electronics – are theyconsistent?” The International Journal of Life Cycle Assessment, vol. 15, no. 8, pp. 827–836,2010. → pages 1, 20, 39, 59, 74[20] M. A. Yao, T. G. Higgs, M. J. Cullen, S. Stewart, and T. A. Brady, “Comparative assessment oflife cycle assessment methods used for personal computers,” Environmental Science &Technology, vol. 44, no. 19, pp. 7335–7346, Oct. 2010. → pages 1, 39, 49, 59, 136[21] A. Sepu´lveda, M. Schluep, F. G. Renaud, M. Streicher, R. Kuehr, C. Hagelu¨ken, and A. C.Gerecke, “A review of the environmental fate and effects of hazardous substances released fromelectrical and electronic equipments during recycling: Examples from China and India,”Environmental Impact Assessment Review, vol. 30, no. 1, pp. 28–41, Jan. 2010. → pages 1, 45[22] O. Tsydenova and M. Bengtsson, “Environmental and human health risks associated with theend-of-life treatment of electrical and electronic equipment,” Institute for Global EnvironmentalStrategies, 2009. [Online]. Available: http://www.indiaenvironmentportal.org.in/node/283545→ pages 1[23] J. Koomey, “Growth in data center electricity use 2005 to 2010,” Analytics Press, Oakland, CA,2011. [Online]. Available: http://www.analyticspress.com/datacenters.html → pages 1, 17, 18,73, 162[24] E. Masanet, R. Brown, A. Shehabi, J. Koomey, and B. Nordman, “Estimating the energy use andefficiency potential of U.S. data centers,” Proceedings of the IEEE, vol. 99, no. 8, pp. 1440–1453, 2011. → pages 1, 163111[25] J. Malmodin, A. Moberg, D. Lunde´n, G. Finnveden, and N. Lo¨vehagen, “Greenhouse gasemissions and operational electricity use in the ICT and entertainment & media sectors,” Journalof Industrial Ecology, vol. 14, no. 5, pp. 770–790, Oct. 2010. → pages 2, 39, 59, 73, 161[26] US Department of Energy, “Buildings energy data book,” Mar. 2012. [Online]. Available:http://buildingsdatabook.eren.doe.gov/ → pages 2, 73[27] L. Gibson, M. Bein, and G. Plamondon, “Innovating toward a low-carbon Canada: Usingtechnology to transform tomorrow,” World Wildlife Fund Canada, Toronto, ON, Canada, 2008.[Online]. Available: http://www.wwf.ca/newsroom/reports/hitech lowcarbon.cfm → pages 2, 9[28] C. Reimsbach-Kounatze, “Towards green ICT strategies: Assessing policies and programmes onICT and the environment,” OECD, No. 155, 2009. [Online]. Available:http://dx.doi.org/10.1787/222431651031 → pages 2[29] T. Kelly and M. Adolph, “ITU-T initiatives on climate change,” Communications Magazine,IEEE, vol. 46, no. 10, pp. 108–114, 2008. → pages 2[30] The Climate Group, Global e-Sustainability Initiative, and McKinsey&Company, “Smart2020:Enabling the low carbon economy in the information age,” Jun. 2008. [Online]. Available:http://www.smart2020.org/ → pages 2, 9, 10, 39, 108[31] V. Turner, R. Bigliani, and C. Ingle, “Reducing greenhouse gases through intense use ofinformation and communication technology: Part 1,” IDC, Dec. 2009. [Online]. Available:http://download.intel.com/pressroom/archive/reference/IDCWP31R.pdf → pages 2[32] M. Griffiths, “ICT and CO2 emissions,” Parliamentary Office of Science and Technology,United Kingdom, 2008. [Online]. Available:http://www.parliament.uk/documents/post/postpn319.pdf → pages 2[33] K. Mallon, G. Johnston, D. Burton, and J. Cavanagh, “Towards a high-bandwidth, low-carbonfuture: Telecommunications-based opportunities to reduce greenhouse gas emissions,” Telstra,2007. [Online]. Available: http://www.telstra.com.au/abouttelstra/download/document/telecommunications-climate-change-blueprint-in-brief.pdf → pages 2, 9[34] US EPA, “ENERGY STAR certified products,” 2014. [Online]. Available:https://data.energystar.gov/ → pages 2[35] J. Katz, W. Rifer, and A. Wilson, “EPEAT: electronic product environmental tool - developmentof an environmental rating system of electronic products for governmental/institutionalprocurement,” in Proceedings of the 2005 IEEE International Symposium on Electronics and theEnvironment, 2005, pp. 1–6. → pages 2, 36[36] L. Deng, E. Williams, and C. Babbitt, “Hybrid life cycle assessment of energy use in laptopcomputer manufacturing,” in 2009 IEEE International Symposium on Sustainable Systems andTechnology ISSST, 2009. → pages 3[37] M. Zgola, E. Olivetti, C. Weber, S. Boyd, J. Mangold, R. Abedrabbo, E. Williams, J. Gregory,and R. Kirchain, “Environmental assessment of information technology products using a triageapproach,” in 2011 IEEE International Symposium on Sustainable Systems and Technology(ISSST), May 2011. → pages 3112[38] A. Moberg, C. Borggren, and G. Finnveden, “Books from an environmental perspective - Part 2:e-books as an alternative to paper books,” The International Journal of Life Cycle Assessment,vol. 16, no. 3, pp. 238–246, 2011. → pages 3, 27, 74[39] C. L. Weber, J. G. Koomey, and H. S. Matthews, “The energy and climate change implicationsof different music delivery methods,” Journal of Industrial Ecology, vol. 14, no. 5, pp. 754–769,2010. → pages 3, 27, 29, 74, 83[40] A. Seetharam, M. Somasundaram, J. Kurose, D. Towsley, and P. Shenoy, “Shipping tostreaming: Is this shift green?” in Proceedings of the first ACM SIGCOMM workshop on Greennetworking, New Delhi, India, 2010. → pages 3, 27, 74[41] D. Schien, P. Shabajee, M. Yearworth, and C. Preist, “Modeling and assessing variability inenergy consumption during the use stage of online multimedia services,” Journal of IndustrialEcology, vol. 17, no. 6, pp. 800–813, 2013. → pages 3, 27, 29, 38, 74, 80, 84, 90, 100, 158, 160,162[42] E. Masanet, A. Shehabi, L. Ramakrishnan, J. Liang, X. Ma, B. Walker, V. Hendrix, andP. Mantha, “The energy efficiency potential of cloud-based software: A U.S. case study,”Lawrence Berkeley National Laboratory, 2013. [Online]. Available:http://crd.lbl.gov/assets/pubs presos/ACS/cloud efficiency study.pdf → pages 3, 27, 74, 82,106, 156, 162, 163[43] Centre for Energy Efficient Telecommunications, “The power of wireless cloud,” University ofMelbourne, 2013. [Online]. Available:http://www.ceet.unimelb.edu.au/pdfs/ceet white paper wireless cloud jun13.pdf → pages 3, 74,158, 160, 162[44] D. Ong, T. Moors, and V. Sivaraman, “Complete life-cycle assessment of the energy/CO2 costsof videoconferencing vs face-to-face meetings,” in 2012 IEEE Online Conference on GreenCommunications (GreenCom), 2012. → pages 3, 12, 27, 74[45] J. Chandaria, J. Hunter, and A. Williams, “A comparison of the carbon footprint of digitalterrestrial television with video-on-demand,” BBC, London, Mar. 2011, WHP 189. → pages 3,27, 29, 38, 74, 95, 162[46] Apple, “Product environmental reports,” 2011. [Online]. Available:http://www.apple.com/environment/reports → pages 6, 60, 66[47] L. Erdmann and L. M. Hilty, “Exploring the macroeconomic impacts of information andcommunication technologies on greenhouse gas emissions,” Journal of Industrial Ecology,vol. 14, no. 5, pp. 826–843, 2010. → pages 8, 9[48] F. Berkhout and J. Hertin, “Impacts of information and communication technologies onenvironmental sustainability: speculations and evidence,” OECD, 2001. → pages 9[49] Forum for the Future, “The impact of ICT on sustainable development,” European InformationTechnology Observatory, 2002. → pages 9[50] L. Erdmann, L. Hilty, J. Goodman, P. Arnfalk, C. Rodriguez Casal, C. Van Wunnik,L. Delgado Sancho, J. Burgelman, and P. Desruelle, “The future impact of ICTs onenvironmental sustainability,” European Commission Joint Research Centre – Institute forProspective Technological Studies, Brussels/Luxembourg, 2004, EUR 21384 EN. → pages 9113[51] M. Buttazzoni, “The potential global CO2 reductions from ICT use: Identifying and assessingthe opportunities to reduce the first billion tonnes of CO2,” WWF Sweden, 2008. [Online].Available: http://www.wwf.se/source.php/1183710/identifying the 1st billion tonnes ict.pdf →pages 9[52] D. MacLean and B. St. Arnaud, “ICTs, innovation and the challenge of climate change,”International Institute for Sustainable Development, Winnipeg, MB, Canada, 2008. [Online].Available: http://site.ebrary.com/lib/ubc/docDetail.action?docID=10232000 → pages 9[53] Bio Intelligence Service, “Impacts of information and communication technologies on energyefficiency,” European Commission, 2008. → pages 9[54] L. Yi and H. R. Thomas, “A review of research on the environmental impact of e-business andICT,” Environment International, vol. 33, no. 6, pp. 841–849, 2007. → pages 10[55] M. Kuhndt, J. von Geibler, V. Tu¨rk, S. Moll, K. O. Schallabo¨ck, and S. Steger, “Virtualdematerialization: ebusiness and factor X,” Wuppertal Institute, Wuppertal, Germany, Mar.2003. → pages 10, 12[56] M. R. Chertow, “The IPAT equation and its variants,” Journal of Industrial Ecology, vol. 4,no. 4, pp. 13–29, 2000. → pages 10[57] M. Draca, R. Sadun, and J. Van Reenen, “Productivity and ICT: a review of the evidence,”Centre for Economic Performance, London School of Economics and Political Science, Aug.2006, discussion Paper 749. → pages 12[58] D. Martı´nez, J. Rodrı´guez, and J. L. Torres, “ICT-specific technological change and productivitygrowth in the US: 1980-2004,” Information Economics and Policy, vol. 22, no. 2, pp. 121–129,May 2010. → pages 12[59] D. W. Jorgenson, M. S. Ho, and K. J. Stiroh, Productivity, Volume 3: Information Technologyand the American Growth Resurgence. The MIT Press, 2005, vol. 3. → pages 12[60] J. Romm, A. Rosenfeld, and S. Herrmann, “The internet economy and global warming: Ascenario of the impact of e-commerce on energy and the environment,” The Center for Energyand Climate Solutions, 1999. → pages 12[61] J. Laitner, J. G. Koomey, E. Worrell, and E. Gumerman, “Re-estimating the annual energyoutlook 2000 forecast using updated assumptions about the information economy,” in AmericanEconomics Association annual meeting, 2001. → pages 12[62] D. Bell, “The coming of the post-industrial society,” The Educational Forum, vol. 40, no. 4, p.574, 1976. → pages 12[63] M. Castells, The Rise of the Network Society, The Information Age: Economy, Society andCulture Vol. I. Cambridge, MA; Oxford, UK: Blackwell, 1996, vol. 12. → pages 12[64] A. S. Duff, Information society studies. Routledge, 2000. → pages 12[65] T. Standage, The Victorian Internet: The remarkable story of the telegraph and the nineteenthcentury’s online pioneers. London: Weidenfeld & Nicolson, 1998. → pages 12[66] I. Salomon, “Telecommunications and travel relationships: a review,” Transportation ResearchPart A: General, vol. 20, no. 3, pp. 223–238, May 1986. → pages 12114[67] E. D. Williams, “Assessing the potential of telecommuting as an energy savings technology injapan,” in Proceedings of the 2003 IEEE International Symposium on Electronics and theEnvironment. IEEE Computer Society, 2003, pp. 147–152. → pages 12[68] E. Kitou and A. Horvath, “Energy-related emissions from telework,” Environmental Science &Technology, vol. 37, no. 16, pp. 3467–3475, 2003. → pages 12, 74[69] P. Mokhtarian, “If telecommunication is such a good substitute for travel, why does congestioncontinue to get worse?” Transportation Letters: The International Journal of TransportationResearch, vol. 1, no. 1, pp. 1–17, Jan. 2009. → pages 12[70] P. L. Mokhtarian, “Telecommunications and travel: The case for complementarity,” Journal ofIndustrial Ecology, vol. 6, no. 2, pp. 43–57, 2002. → pages 12[71] S. Choo and P. L. Mokhtarian, “Telecommunications and travel demand and supply: Aggregatestructural equation models for the US,” Transportation Research Part A: Policy and Practice,vol. 41, no. 1, pp. 4–18, Jan. 2007. → pages 12[72] V. C. Coroama and L. M. Hilty, “Assessing internet energy intensity: A review of methods andresults,” Environmental Impact Assessment Review, vol. 45, pp. 63–68, Feb. 2014. → pages 14,24, 82, 84, 161[73] IPCC, “Contribution of working groups I, II and III to the fourth assessment report of theintergovernmental panel on climate change,” Geneva, Switzerland, 2007. → pages 15[74] J. C. Bare, “Traci: The tool for the reduction and assessment of chemical and otherenvironmental impacts,” Journal of Industrial Ecology, vol. 6, no. 3-4, pp. 49–78, 2002. →pages 15[75] O. Jolliet, M. Margni, R. Charles, S. Humbert, J. Payet, G. Rebitzer, and R. Rosenbaum,“IMPACT 2002+: A new life cycle impact assessment methodology,” The International Journalof Life Cycle Assessment, vol. 8, no. 6, pp. 324–330, Nov. 2003. → pages 15[76] C. Bulle, M. Margini, O. Jolliet, R. Rosenbaum, and S. Humbert, “Impact world+,” 2012.[Online]. Available: http://www.impactworldplus.org/ → pages 15[77] M. Hauschild and J. Potting, “Spatial differentiation in life cycle impact assessment - theEDIP2003 methodology,” Danish Environmental Protection Agency, 2005. → pages 15[78] J. B. Guine´e, M. Gorre´e, R. Heijungs, G. Huppes, R. Kleijn, A. de Koning, L. van Oers,A. Wegener Sleeswijk, S. Suh, H. A. Udo de Haes et al., “Life cycle assessment; an operationalguide to the ISO standards.” Ministry of Housing, Spatial Planning and Environment (VROM)and Centre of Environmental Science (CML), Den Haag and Leiden, The Netherlands, 2001. →pages 15[79] G. Finnveden, M. Z. Hauschild, T. Ekvall, J. Guine, R. Heijungs, S. Hellweg, A. Koehler,D. Pennington, and S. Suh, “Recent developments in life cycle assessment,” Journal ofEnvironmental Management, vol. 91, no. 1, pp. 1–21, Oct. 2009. → pages 15, 34, 37[80] L. B. Lave, E. Cobas-Flores, C. T. Hendrickson, and F. C. McMichael, “Using input-outputanalysis to estimate economy-wide discharges,” Environmental Science & Technology, vol. 29,no. 9, pp. 420A–426A, Sep. 1995. → pages 15, 34115[81] G. Majeau-Bettez, A. H. Strømman, and E. G. Hertwich, “Evaluation of process- andinput-output-based life cycle inventory data with regard to truncation and aggregation issues,”Environmental Science & Technology, vol. 45, no. 23, pp. 10 170–10 177, 2011. → pages 15, 34,61[82] K. W. Roth, F. Goldstein, and J. Kleinman, “Energy consumption by office andtelecommunications equipment in commercial buildings volume I: Energy consumptionbaseline,” Arthur D. Little, Inc, Cambridge, MA, Jan. 2002, Reference No. 72895-00. → pages16, 17, 50, 51[83] J. Koomey, “Estimating total power consumption by servers in the U.S. and the world.”Analytics Press, Oakland, CA, Feb. 2007. → pages 17, 18[84] Google, “Efficiency: How we do it,” 2014. [Online]. Available:http://www.google.com/about/datacenters/efficiency/internal/#measuring-efficiency → pages 18[85] Facebook, “Prineville, OR data center,” 2014. [Online]. Available:https://www.fbpuewue.com/prineville → pages 18[86] G. Moore, “Cramming more components onto integrated circuits,” Electronics Magazine, 1965.→ pages 19[87] R. Keyes, “The impact of Moore’s Law,” IEEE Solid-State Circuits Society Newsletter, vol. 11,no. 5, pp. 25–27, Sep. 2006. → pages 19[88] ITRS, “International technology roadmap for semiconductors, 2013 edition, executivesummary.” ITRS, 2013. [Online]. Available: www.itrs.net → pages 19[89] M. Bohr, “A 30 year retrospective on Dennard’s MOSFET scaling paper,” IEEE Solid-StateCircuits Society Newsletter, vol. 12, no. 1, pp. 11–13, 2007. → pages 19[90] J. Koomey, S. Berard, M. Sanchez, and H. Wong, “Implications of historical trends in theelectrical efficiency of computing,” IEEE Annals of the History of Computing, vol. 33, no. 3, pp.46–54, Mar. 2011. → pages 19[91] Y. Arushanyan, E. Ekener-Petersen, and G. Finnveden, “Lessons learned – review of LCAs forICT products and services,” Computers in Industry, vol. 65, no. 2, pp. 211–234, Feb. 2014. →pages 20, 21, 74[92] L. Deng, C. W. Babbitt, and E. D. Williams, “Economic-balance hybrid LCA extended withuncertainty analysis: case study of a laptop computer,” Journal of Cleaner Production, vol. 19,no. 11, pp. 1198–1206, Jul. 2011. → pages 20, 34, 35, 61[93] A. Andrae and O. Andersen, “Life cycle assessment of integrated circuit packagingtechnologies,” The International Journal of Life Cycle Assessment, vol. 16, no. 3, pp. 258–267,2011. → pages 21, 49, 63[94] R. Frischknecht, N. Jungbluth, H.-J. Althaus, G. Doka, R. Dones, T. Heck, S. Hellweg,R. Hischier, T. Nemecek, G. Rebitzer, and M. Spielmann, “The ecoinvent database: Overviewand methodological framework,” The International Journal of Life Cycle Assessment, vol. 10,no. 1, pp. 3–9, Oct. 2004. → pages 21116[95] R. Hischier, M. Classen, M. Lehmann, and W. Scharnhorst, “Life cycle inventories of electricand electronic equipment: Production, use and disposal,” Swiss Centre for Life CycleInventories, Du¨bendorf, Switzerland, 2007, final report ecoinvent data v2.0 No. 18. → pages 21,44, 47, 51, 53, 59, 60, 61, 62, 134, 135, 146[96] P. Teehan, S. Storey, and M. Kandlikar, “Rapid integrated life cycle assessment of building-wideIT systems,” in 2011 IEEE International Symposium on Sustainable Systems and Technology(ISSST), May 2011. → pages 21[97] J. Koomey, H. Chong, W. Loh, B. Nordman, and M. Blazek, “Network electricity use associatedwith wireless personal digital assistants,” Journal of Infrastructure Systems, vol. 10, no. 3, pp.131–137, 2004. → pages 24, 82[98] Nielsen, “Free to move between screens: The cross-platform report,” Mar. 2013. [Online].Available: http://www.nielsen.com/content/dam/corporate/us/en/reports-downloads/2013%20Reports/Nielsen-March-2013-Cross-Platform-Report.pdf → pages 25, 73, 86, 166, 167[99] Digital TV Research, “Connected TV forecasts report,” Oct. 2013. [Online]. Available:https://www.digitaltvresearch.com/ → pages 25[100] Alcatel-Lucent, “Video shakes up the IP edge,” Dec. 2012. [Online]. Available:http://resources.alcatel-lucent.com/?cid=162848 → pages 25, 26, 96, 166[101] comScore, Inc., “US digital future in focus: 2013,” 2013. [Online]. Available: http://www.comscore.com/Insights/Presentations and Whitepapers/2013/2013 US Digital Future in Focus→ pages 25, 166, 167[102] Cisco Systems Inc., “Cisco visual networking index: Forecast and methodology, 2012-2017,”2013. [Online]. Available: http://www.cisco.com/c/en/us/solutions/collateral/service-provider/ip-ngn-ip-next-generation-network/white paper c11-481360.html → pages 25, 157, 158, 159,162, 163, 169[103] ——, “VNI mobile forecast highlights, 2012 - 2017,” 2013. [Online]. Available:http://www.cisco.com/web/solutions/sp/vni/vni mobile forecast highlights/index.html → pages25, 75, 82, 84, 86, 97, 163[104] comScore, Inc., “Mobile future in focus 2013,” 2013. [Online]. Available: https://www.comscore.com/Insights/Presentations and Whitepapers/2013/2013 Mobile Future in Focus3 →pages 25[105] Ericsson, “Ericsson mobility report: On the pulse of the networked society,” Jun. 2013. [Online].Available: http://www.ericsson.com/mobility-report → pages 25[106] M. Missbach, J. Stelzel, C. Gardiner, G. Anderson, and M. Tempes, “A short history of cloudcomputing,” in SAP on the Cloud. Springer Berlin Heidelberg, Jan. 2013, pp. 1–13. → pages25[107] P. Mell and T. Grance, “The NIST definition of cloud computing,” National Institute ofStandards and Technology, Gaithersburg, MD, 2011, NIST Special Publication 800-145.[Online]. Available: http://csrc.nist.gov/publications/nistpubs/800-145/SP800-145.pdf → pages25117[108] Cisco Systems Inc., “Cisco global cloud index: Forecast and methodology, 2011–2016,” 2012.[Online]. Available:http://www.cisco.com/en/US/netsol/ns1175/networking solutions solution category.html →pages 26, 162[109] T. Danova, “SMART WATCH FORECAST: smart watches will lead wearable computing tomass market status,” Business Insider, Mar. 2014. [Online]. Available:http://www.businessinsider.com/global-smartwatch-sales-set-to-explode-2014-3 → pages 26[110] J. Happich, “Spurred by Google Glass, IHS forecasts nearly 10 million smart glasses to shipfrom 2012 to 2016,” EETimes, May 2013. [Online]. Available:http://www.eetimes.com/document.asp?doc id=1266624 → pages 26[111] V. Sehgal and S. R. Epps, “Forrester research consumer PC and tablet forecast, 2011 to 2016(US).” Forrester Research, Inc., Jan. 2012. → pages 26, 82, 157[112] eMarketer, “Smartphones in the hands of the youngest demographic,” May 2013. [Online].Available:http://www.emarketer.com/Article/Smartphones-Hands-of-Youngest-Demographic/1009915 →pages 26, 82, 86, 157[113] G. Doyle, “From television to multi-platform: Less from more or more for less?” Convergence:The International Journal of Research into New Media Technologies, vol. 16, no. 4, pp.431–449, Nov. 2010. → pages 26[114] IDC, “IDC forecasts worldwide public IT cloud services spending to reach nearly $108 billionby 2017 as focus shifts from savings to innovation,” Sep. 2013. [Online]. Available:http://www.idc.com/getdoc.jsp?containerId=prUS24298013 → pages 26[115] S. Florentine and T. Olavsrud, “2014 forecast for cloud computing,” CIO, Dec. 2013. [Online].Available: http://www.cio.com/article/745155/2014 Forecast for Cloud Computing → pages26[116] J. G. Bull and R. A. Kozak, “Comparative life cycle assessments: The case of paper and digitalmedia,” Environmental Impact Assessment Review, vol. 45, pp. 10–18, Feb. 2014. → pages 26[117] M. F. Emmenegger, R. Frischknecht, M. Stutz, M. Guggisberg, R. Witschi, and T. Otto, “Lifecycle assessment of the mobile communication system UMTS,” International Journal of LIfeCycle Assessment, 2004. → pages 27[118] M. W. Toffel and A. Horvath, “Environmental implications of wireless technologies: Newsdelivery and business meetings,” Environmental Science & Technology, vol. 38, no. 11, pp.2961–2970, Jun. 2004. → pages 27, 74[119] Y. Matsuno, K. Takahashi, and M. Tsuda, “Eco-efficiency for information and communicationstechnology (ICT): the state of knowledge in Japan,” in Proceedings of the 2007 IEEEInternational Symposium on Electronics and the Environment, 2007, pp. 1–5. → pages 27[120] C. Borggren, A. Moberg, M. Ra¨sa¨nen, and G. Finnveden, “Business meetings at a distance -decreasing greenhouse gas emissions and cumulative energy demand?” Journal of CleanerProduction, vol. 41, pp. 126–139, Feb. 2013. → pages 27118[121] G. L. Kozak and G. A. Keolelan, “Printed scholarly books and e-book reading devices: acomparative life cycle assessment of two book options,” in Proceedings of the 2003 IEEEInternational Symposium on Electronics and the Environment. IEEE Computer Society, 2003,pp. 291–296. → pages 27, 28, 74[122] M. Enroth, “Environmental impact of printed and electronic teaching aids, a screening studyfocusing on fossil carbon dioxide emission,” Advances in Printing and Media Technology,vol. 36, p. 2330, 2009. → pages 27, 28[123] M. Ahmadi Achachlouei, A. Moberg, and E. Hochschorner, “Climate change impact ofelectronic media solutions: Case study of the tablet edition of a magazine,” in ICT4S 2013:Proceedings of the First International Conference on Information and CommunicationTechnologies for Sustainability, 2013. → pages 27[124] D. L. Gard and G. A. Keoleian, “Digital versus print: Energy performance in the selection anduse of scholarly journals,” Journal of Industrial Ecology, vol. 6, no. 2, pp. 115–132, 2002. →pages 27, 74[125] L. Farrant and Y. Le Guern, “Which environmental impacts for ICT? – LCA case study onelectronic mail,” in Electronics Goes Green 2012+. IEEE, 2012, p. 15. → pages 27[126] I. Reichart, “The environmental impact of getting the news,” Journal of Industrial Ecology,vol. 6, no. 3-4, pp. 185–200, 2002. → pages 27, 74[127] A. Moberg, M. Johansson, G. Finnveden, and A. Jonsson, “Printed and tablet e-paper newspaperfrom an environmental perspective a screening life cycle assessment,” Environmental ImpactAssessment Review, vol. 30, no. 3, pp. 177–191, Apr. 2010. → pages 27[128] C. Taylor and J. Koomey, “Estimating energy use and greenhouse gas emissions of internetadvertising.” Feb. 2008. [Online]. Available: http://imc2.com/Documents/CarbonEmissions.pdf→ pages 27, 82[129] D. R. Williams and Y. Tang, “Impact of office productivity cloud computing on energyconsumption and greenhouse gas emissions,” Environmental Science & Technology, vol. 47,no. 9, pp. 4333–4340, May 2013. → pages 27, 74[130] D. Sivaraman, S. Pacca, K. Mueller, and J. Lin, “Comparative energy, environmental, andeconomic analysis of traditional and e-commerce DVD rental networks,” Journal of IndustrialEcology, vol. 11, no. 3, pp. 77–91, 2007. → pages 27, 74[131] J. Roettgers, “The slow but inevitable decline of Netflix’s DVD business,” GigaOm, Oct. 2013.[Online]. Available:http://gigaom.com/2013/10/21/when-will-netflix-kill-its-dvd-subscriptions/ → pages 28[132] D. Schien, P. Shabajee, S. G. Wood, and C. Preist, “A model for green design of online newsmedia services,” in Proceedings of the 22nd international conference on World Wide Web,Geneva, 2013, pp. 1111–1122. → pages 29, 100[133] C. L. Weber, “Uncertainty and variability in product carbon footprinting,” Journal of IndustrialEcology, vol. 16, no. 2, pp. 203–211, Apr. 2012. → pages 33, 62, 65, 141, 169119[134] J. Reap, F. Roman, S. Duncan, and B. Bras, “A survey of unresolved problems in life cycleassessment. Part 1: goal and scope and inventory analysis,” The International Journal of LifeCycle Assessment, vol. 13, no. 4, pp. 290–300, Jun. 2008. → pages 33[135] ——, “A survey of unresolved problems in life cycle assessment. Part 2: impact assessment andinterpretation,” The International Journal of Life Cycle Assessment, vol. 13, no. 5, pp. 374–388,2008. → pages 33[136] C. Hendrickson, A. Horvath, S. Joshi, and L. Lave, “Economic input-output models forenvironmental life-cycle assessment,” Environmental science & technology (policy analysis),vol. 32, no. 7, pp. 184A–191A, 1998. → pages 34[137] S. Suh, M. Lenzen, G. J. Treloar, H. Hondo, A. Horvath, G. Huppes, O. Jolliet, U. Klann,W. Krewitt, Y. Moriguchi, J. Munksgaard, and G. Norris, “System boundary selection inlife-cycle inventories using hybrid approaches,” Environmental Science & Technology, vol. 38,no. 3, pp. 657–664, Feb. 2004. → pages 34[138] Y. Loerincik, “Environmental impacts and benefits of information and communicationtechnology infrastructure and services, using process and input-output life cycle assessment,”Ph.D. dissertation, E´cole Polytechnique Fe´de´rale de Lausanne, Lausanne, Switzerland, 2006. →pages 34, 35, 61[139] S. Suh and G. Huppes, “Methods for life cycle inventory of a product,” Journal of CleanerProduction, vol. 13, no. 7, pp. 687–697, Jun. 2005. → pages 34, 61[140] M. Lenzen, “Differential convergence of life-cycle inventories toward upstream productionlayers,” Journal of Industrial Ecology, vol. 6, no. 3-4, pp. 137–160, Jul. 2002. → pages 34[141] R. K. Rosenbaum, T. M. Bachmann, L. S. Gold, M. A. J. Huijbregts, O. Jolliet, R. Juraske,A. Koehler, H. F. Larsen, M. MacLeod, M. Margni, T. E. McKone, J. Payet, M. Schuhmacher,D. v. d. Meent, and M. Z. Hauschild, “USEtox – the UNEP-SETAC toxicity model:recommended characterisation factors for human toxicity and freshwater ecotoxicity in lifecycle impact assessment,” The International Journal of Life Cycle Assessment, vol. 13, no. 7, pp.532–546, Nov. 2008. → pages 36[142] A. O. W. Leung, N. S. Duzgoren-Aydin, K. C. Cheung, and M. H. Wong, “Heavy metalsconcentrations of surface dust from e-waste recycling and its human health implications insoutheast China,” Environmental Science & Technology, vol. 42, no. 7, pp. 2674–2680, Apr.2008. → pages 36[143] J. Fu, Q. Zhou, J. Liu, W. Liu, T. Wang, Q. Zhang, and G. Jiang, “High levels of heavy metals inrice (oryza sativa l.) from a typical e-waste recycling area in southeast China and its potentialrisk to human health,” Chemosphere, vol. 71, no. 7, pp. 1269–1275, Apr. 2008. → pages 36[144] Q. Luo, Z. W. Cai, and M. H. Wong, “Polybrominated diphenyl ethers in fish and sediment fromriver polluted by electronic waste,” Science of The Total Environment, vol. 383, no. 1-3, pp.115–127, Sep. 2007. → pages 36[145] M. Wong, S. Wu, W. Deng, X. Yu, Q. Luo, A. Leung, C. Wong, W. Luksemburg, and A. Wong,“Export of toxic chemicals - a review of the case of uncontrolled electronic-waste recycling,”Environmental Pollution, vol. 149, no. 2, pp. 131–140, Sep. 2007. → pages 36120[146] W. Deng, J. Zheng, X. Bi, J. Fu, and M. Wong, “Distribution of PBDEs in air particles from anelectronic waste recycling site compared with Guangzhou and Hong Kong, South China,”Environment International, vol. 33, no. 8, pp. 1063–1069, Nov. 2007. → pages 36[147] X. Huo, L. Peng, X. Xu, L. Zheng, B. Qiu, Z. Qi, B. Zhang, D. Han, and Z. Piao, “Elevatedblood lead levels of children in Guiyu, an electronic waste recycling town in China,”Environmental Health Perspectives, 2007. → pages 36[148] L. Zheng, K. Wu, Y. Li, Z. Qi, D. Han, B. Zhang, C. Gu, G. Chen, J. Liu, S. Chen, X. Xu, andX. Huo, “Blood lead and cadmium levels and relevant factors among children from an e-wasterecycling town in China,” Environmental Research, vol. 108, no. 1, pp. 15–20, Sep. 2008,PMID: 18514186. → pages 36[149] Y. Li, X. Xu, J. Liu, K. Wu, C. Gu, G. Shao, S. Chen, G. Chen, and X. Huo, “The hazard ofchromium exposure to neonates in Guiyu of China,” Science of The Total Environment, vol. 403,no. 1-3, pp. 99–104, Sep. 2008. → pages 36[150] Y. Li, X. Xu, K. Wu, G. Chen, J. Liu, S. Chen, C. Gu, B. Zhang, L. Zheng, M. Zheng, andX. Huo, “Monitoring of lead load and its effect on neonatal behavioral neurological assessmentscores in Guiyu, an electronic waste recycling town in China,” Journal of EnvironmentalMonitoring: JEM, vol. 10, no. 10, pp. 1233–1238, Oct. 2008, PMID: 19244648. → pages 36[151] European Union, “Directive 2002/95/EC of the European Parliament and of the Council of 27January 2003 on the restriction of the use of certain hazardous substances in electrical andelectronic equipment (RoHS),” 2002. → pages 36[152] ——, “Directive 2002/96/EC of the European parliament and of the council of 27 January 2003on waste electrical and electronic equipment (WEEE) joint declaration of the Europeanparliament, the council and the commission relating to article 9,” 2002. → pages 36[153] M. Wackernagel and W. Rees, Our Ecological Footprint: Reducing Human Impact on the Earth.New Society Publishers, Jul. 1998. → pages 37[154] J. Spangenberg, F. Hinterberger, S. Moll, and H. Schutz, “Material flow analysis, TMR and theMIPS concept: a contribution to the development of indicators for measuring changes inconsumption and production patterns,” International Journal of Sustainable Development,vol. 2, no. 4, pp. 491–505, Jan. 1999. → pages 37[155] M. Lenzen, “Structural path analysis of ecosystem networks,” Ecological Modelling, vol. 200,no. 34, pp. 334–342, Jan. 2007. → pages 37[156] P. James and L. Hopkinson, “Energy and environmental impacts of personal computing: A bestpractice review prepared for the joint information services committee (JISC),” SustainIT, UKCentre for Economic and Environmental Development, May 2009. → pages 39[157] MCC, “Environmental consciousness: A strategic competitiveness issue for the electronics andcomputer industry, summary report: International analysis, conclusions and recommendations,”Microelectronics and Computer Technology Corporation, Austin, TX, Mar. 1993. → pages 42[158] R. Eugster, R. Hischier, and H. Duan, “Key environmental impacts of the Chinese EEE industry– a life cycle assessment study,” 2007. → pages 42, 43, 44121[159] H. Duan, M. Eugster, R. Hischier, M. Streicher-Porte, and J. Li, “Life cycle assessment study ofa Chinese desktop personal computer,” Science of The Total Environment, vol. 407, no. 5, pp.1755–1764, Feb. 2009. → pages 42, 43, 44, 51, 53, 61, 135[160] M. Tekawa, S. Miyamoto, and A. Inaba, “Life cycle assessment; an approach to environmentallyfriendly PCs,” in Proceedings of the 1997 IEEE International Symposium on Electronics and theEnvironment, 1997, pp. 125–130. → pages 42, 44, 53, 134[161] Atlantic Consulting, “LCA study of the product group personal computers in the EU Ecolabelscheme,” Ecolabel Unit of the European Commission, 1998. → pages 43, 51, 53, 55[162] E. Masanet, L. Price, S. de la Rue du Can, R. Brown, and E. Worrell, “Optimization of productlife cycles to reduce greenhouse gas emissions in california,” LBNL and CEC, 2005,LBNL-57029 and CEC-500-2005-110-FX. → pages 43, 45, 53, 55, 56[163] E. Williams and Y. Sasaki, “Energy analysis of end-of-life options for personal computers:resell, upgrade, recycle,” in Proceedings of the 2003 IEEE International Symposium onElectronics and the Environment, 2003, pp. 187–192. → pages 43[164] K. Kawamoto, J. Koomey, B. Nordman, R. Brown, M. A. Piette, M. Ting, and A. K. Meier,“Electricity used by office equipment and network equipment in the U.S.: detailed report andappendices,” Lawrence Berkeley National Laboratory, Berkeley, CA, 2001, LBNL-45917. →pages 43, 50, 51[165] B. P. Hikwama, “Life cycle assessment of a personal computer,” Research Thesis, University ofSouthern Queensland, 2005. → pages 43, 44, 47, 51, 53, 134[166] R. Kemna, “Methodology study of energy-using products,” VHK, European Commission, 2005.→ pages 43, 44, 50, 51, 53, 61[167] B.-C. Choi, H.-S. Shin, S.-Y. Lee, and T. Hur, “Life cycle assessment of a personal computerand its effective recycling rate,” The International Journal of Life Cycle Assessment, vol. 11,no. 2, pp. 122–128, Mar. 2006. → pages 43, 45, 51, 52, 53, 98[168] A. Braune and M. Held, “Development of environmental performance indicators for ICTproducts on the example of personal computer,” European Commission, 2006. → pages 43, 44,51, 53, 61, 134[169] IVF, “Preparatory studies for eco-design requirements of EuPs, lot 3, personal computres(desktops and laptops) and computer monitors,” IVF Industrial Research and DevelopmentCorporation, 2007. [Online]. Available: http://www.ecocomputer.org → pages 44, 47, 51, 53,61, 134[170] Apple, “Mac mini environmental report,” 2010. [Online]. Available:http://www.apple.com/environment/reports/ → pages 44, 55, 56[171] ——, “Mac pro environmental report,” 2010. [Online]. Available:http://www.apple.com/environment/reports/ → pages 44, 55, 56[172] S. B. Boyd, A. Horvath, and D. Dornfeld, “Life-cycle energy demand and global warmingpotential of computational logic,” Environmental Science & Technology, vol. 43, no. 19, pp.7303–7309, Oct. 2009. → pages 49, 62, 135, 141122[173] US EPA, “Savings calculator for ENERGY STAR qualified office equipment,” US EPA, 2011.[Online]. Available: https://www.energystar.gov/ia/products/fap//Calc office eq.xls → pages 51[174] ——, “ENERGY STAR computer specification energy savings assumptions,” US EPA, Report,2005. → pages 51[175] L. Moorefield, B. Frazer, and P. Bendt, “Office plug load field monitoring report: DRAFT,”Ecos consulting, Durango, CO, Dec. 2008. → pages 51[176] S. F. Porter, L. Moorefield, and P. May-Ostendorp, “Final field research report,” EcosConsulting, Oct. 2006. → pages 51[177] MTP, “Monitoring home computers,” Market Transformation Program, UK, Report, 2006. →pages 51[178] B. Schlomann, C. Cremer, M. Friedewald, P. Georgieff, E. Gruber, R. Corrandini, D. Kraus,U. Arndt, W. Mauch, H. Schaefer et al., “Technical and legal application possibilities of thecompulsory labeling of the standby consumption of electrical household and office appliances,”Fraunhofer ISI., 2005, summary of the Final Report for the Federal Ministry of Economics andLabour, Project No. 53/03. → pages 51[179] B. Nordman and A. Meier, “Energy consumption of home information technology,” LawrenceBerkeley National Laboratory, 2004, LBNL-5350. → pages 51[180] C. Babbitt, R. Kahhat, E. Williams, and G. Barbitt, “Evolution of product lifespan andimplications for environmental assessment and management: A case study of personalcomputers in higher education.” Environmental Science & Technology, vol. 43, no. 13, pp.5106–5112, Jul. 2009. → pages 52, 53[181] E. Williams and T. Hatanaka, “Residential computer usage patterns in japan and associated lifecycle energy use,” in Proceedings of the 2005 International Symposium on Electronics and theEnvironment. IEEE Computer Society, 2005, pp. 177–182. → pages 52, 53[182] C. Smulders, “Watching rome burn: PC empires threatened by extending life cycles,” Gartner,2001. [Online]. Available:https://www.gartner.com/doc/338985/watching-rome-burn-pc-empires → pages 52, 53[183] JEITA, “Research report on collection, disposal, recycling situation of IT products (inJapanese),” Japan Electronics and Information Technology Industries Association, 2006. →pages 52, 53[184] M. Oguchi, M. Kameya, and T. Tasaki, “Estimation of lifetime distributions and waste numbersof 23 types of electrical and electronic equipment (in Japanese),” Journal of Japan WasteManagement Association, vol. 17, pp. 50–60, 2006. → pages 52, 53[185] ESRI, “Monthly consumer confidence survey covering all of Japan (summary),” Economic andSocial Resaerch Institute, 2007. → pages 52, 53[186] A. Yoshida, T. Tasaki, and A. Terazono, “Material flow of used PCs in japan,” in Proceedings ofthe 2007 IEEE International Symposium on Electronics and the Environment, 2007, pp. 46–51.→ pages 52, 53123[187] ——, “Material flow analysis of used personal computers in Japan,” Waste Management,vol. 29, no. 5, pp. 1602–1614, May 2009. → pages 52, 53[188] J. Fiskel, Design for environment: creating eco-efficient products and processes.McGraw-Hill, 1996. → pages 53[189] R. Dones, C. Bauer, R. Bolliger, B. Burger, M. Faist Emmenegger, R. Frischknecht, T. Heck,N. Jungbluth, A. Roder, and M. Tuchschmid, “Life cycle inventories of energy systems: resultsfor current systems in switzerland and other UCTE countries.” Paul Scherrer Institut Villigen,Swiss Centre for Life Cycle Inventories, Dubendorf, Switzerland, 2007. → pages 54[190] E. G. Hertwich and C. Roux, “Greenhouse gas emissions from the consumption of electric andelectronic equipment by norwegian households,” Environmental Science& Technology, vol. 45,no. 19, pp. 8190–8196, 2011. → pages 59[191] R. Frischknecht, N. Jungbluth, H.-J. Althaus, G. Doka, R. Dones, T. Heck, S. Hellweg,R. Hischier, T. Nemecek, G. Rebitzer, and M. Spielmann, “The ecoinvent database: Overviewand methodological framework,” The International Journal of Life Cycle Assessment, vol. 10,no. 1, pp. 3–9, Oct. 2004. → pages 59, 67, 140, 143, 144[192] MIT Material Systems Laboratory, “Product attribute to impact algorithm - PAIA,” Jun. 2012.[Online]. Available: http://msl.mit.edu/index.php?id=33 → pages 60[193] T. Okrasinski and J. Malian, “A framework for estimating life-cycle eco-impact for informationand communications technology products,” in CARE Conference, Vienna, 2010. → pages 60[194] ETSI, “Environmental engineering (EE); life cycle assessment (LCA) of ICT equipment,networks and services; general methodology and common requirements,” EuropeanTelecommunications Standards Institute, Nov. 2011, ETSI TS 103 199 V1.1.1. [Online].Available:http://www.etsi.org/deliver/etsi ts/103100 103199/103199/01.01.01 60/ts 103199v010101p.pdf→ pages 62, 72[195] ITU, “Methodology for energy consumption and greenhouse gas emissions impact assessmentof information and communication technologies in organizations,” InternationalTelecommunication Union, ITU-T L1420. [Online]. Available:http://www.itu.int/rec/T-REC-L.1420/en → pages 62[196] IEC, “Quantification methodology of greenhouse gas emissions (CO2e) for electrical andelectronic products and systems,” International Electrotechnical Commission, 2012, IEC TC 111- IEC/TR 62725. → pages 62[197] GHG Protocol, “GHG protocol product life cycle accounting and reporting standard ICT sectorguidance (draft),” Mar. 2012. [Online]. Available:http://www.ghgprotocol.org/standards/product-standard → pages 62[198] European Commission DG Information Society, “ICT for sustainable growth.” [Online].Available:http://ec.europa.eu/information society/activities/sustainable growth/ict sector/index en.htm →pages 62124[199] M. Stutz, “Product carbon footprint (PCF) assessment of a Dell OptiPlex 780 desktop - resultsand recommendations,” in Towards Life Cycle Sustainability Management, M. Finkbeiner, Ed.Netherlands: Springer, 2011, vol. 8, pp. 495–500. → pages 65[200] S. O’Connell and M. Stutz, “Product carbon footprint (PCF) assessment of a Dell laptop -results and recommendations,” in 2010 IEEE International Symposium on Sustainable Systemsand Technology (ISSST), May 2010, pp. 1 –6. → pages 65[201] C. Borggren and A. Moberg, “Pappersbok och elektronisk bok pa˚ la¨splatta,” KTH Centre forSustainable Communications, Stockholm, Sweden, 2009. → pages 66[202] F. Bergelin, “Life cycle assessment of a mobile phone: a model on manufacturing, using andrecycling.” Master’s thesis, Uppsala University, Uppsala, Sweden, 2008. → pages 66[203] PE International, “Environmental footprint of ICT equipment in manufacture, use, andend-of-life.” in European Conference on Optical Communication, Brussels, Belgium, Sep. 2008.→ pages 66[204] M. G. Morgan, M. Henrion, and M. Small, Uncertainty: a guide to dealing with uncertainty inquantitative risk and policy analysis. Cambridge University Press, 1990. → pages 67[205] M. Zgola, “A triage approach to streamline environmental footprinting : a case study for liquidcrystal displays,” MSc thesis, Massachusetts Institute of Technology, Cambridge, MA, 2011. →pages 71[206] K. Roth and K. McKenney, “Residential consumer electronics electricity consumption in theUnited States,” in ECEEE 2007 Summer Study, 2007. → pages 73[207] US EIA, “Annual energy outlook 2013 with projections to 2040,” US Dept of Energy, EnergyInformation Administration, 2013. [Online]. Available: http://www.eia.gov/forecasts/aeo →pages 73[208] S. Lanzisera, B. Nordman, and R. E. Brown, “Data network equipment energy use and savingspotential in buildings,” Energy Efficiency, vol. 5, no. 2, pp. 149–162, May 2012. → pages 73,160, 161[209] eMarketer, “How digital time spent breaks down by device, gender, content area,” Sep. 2013.[Online]. Available: http://www.emarketer.com/Article/How-Digital-Time-Spent-Breaks-Down-by-Device-Gender-Content-Area/1010239 → pages73, 166, 167[210] L. Deng, C. W. Babbitt, and E. D. Williams, “Economic-balance hybrid LCA extended withuncertainty analysis: case study of a laptop computer,” Journal of Cleaner Production, vol. 19,no. 11, pp. 1198–1206, Jul. 2011. → pages 74[211] A. Moberg, C. Borggren, C. Ambell, G. Finnveden, F. Guldbrandsson, A. Bondesson,J. Malmodin, and P. Bergmark, “Simplifying a life cycle assessment of a mobile phone,” TheInternational Journal of Life Cycle Assessment, pp. 1–15, 2014. → pages 74[212] Apple, “Product environmental reports,” 2014. [Online]. Available:http://www.apple.com/environment/reports → pages 74, 99, 104125[213] H. Matthews, C. Hendrickson, and D. Soh, “Environmental and economic effects ofe-commerce: A case study of book publishing and retail logistics,” Transportation ResearchRecord: Journal of the Transportation Research Board, vol. 1763, no. -1, pp. 6–12, Jan. 2001.→ pages 74[214] K. Roth, T. Rhodes, and R. Ponoum, “The energy and greenhouse gas emission impacts oftelecommuting in the U.S.” in Proceedings of the 2008 IEEE International Symposium onElectronics and the Environment, 2008. → pages 74[215] P. M. Corcoran and A. S. G. Andrae, “Emerging trends in electricity consumption for consumerICT,” NUI Galway School of Electrical and Electronic Engineering, 2013. [Online]. Available:http://vmserver14.nuigalway.ie/xmlui/handle/10379/3563 → pages 74[216] US EPA, “eGRID 9th edition year 2010,” Feb. 2014. [Online]. Available:http://www.epa.gov/egrid → pages 77[217] EPRI, “EPRI calculates annual cost of charging an iPad at $1.36,” Jun. 2012. [Online].Available: http://www.epri.com/Press-Releases/Pages/EPRI-Calculates-Annual-Cost-of-Charging-an-iPad-at-$1-36.aspx → pages 82, 155[218] IDC, “PC outlook falls as market increasingly looks to tablets, according to IDC,” May 2013.[Online]. Available: http://www.idc.com/getdoc.jsp?containerId=prUS24129913 → pages 82,157[219] J. Baliga, R. Ayre, K. Hinton, and R. Tucker, “Energy consumption in wired and wireless accessnetworks,” IEEE Communications Magazine, vol. 49, no. 6, pp. 70–77, 2011. → pages 83, 160[220] M. Pickavet, W. Vereecken, S. Demeyer, P. Audenaert, B. Vermeulen, C. Develder, D. Colle,B. Dhoedt, and P. Demeester, “Worldwide energy needs for ICT: the rise of power-awarenetworking,” in 2nd International Symposium on Advanced Networks and TelecommunicationSystems, 2008, pp. 1–3. → pages 83[221] V. C. Coroama, L. M. Hilty, E. Heiri, and F. M. Horn, “The direct energy demand of internetdata flows,” Journal of Industrial Ecology, p. n/an/a, 2013. → pages 84, 161[222] comScore, Inc., “2012 Mobile future in focus: Key insights from 2012 and what they mean forthe coming year,” 2012. [Online]. Available: http://www.comscore.com/Insights/Presentations and Whitepapers/2012/2012 Mobile Future in Focus → pages 86[223] G. Cook, T. Dowdall, D. Pomerantz, and Y. Wang, “Clicking clean: How companies are creatingthe green internet,” Greenpeace Inc, Washington DC, Apr. 2014. [Online]. Available:http://www.greenpeace.org/usa/Global/usa/planet3/PDFs/clickingclean.pdf → pages 90, 105[224] Google, “Google green: The big picture.” [Online]. Available:http://www.google.ca/green/bigpicture/ → pages 90, 91, 162[225] US Department of Energy, “Energy saver tips: appliances,” Jul. 2012. [Online]. Available:http://energy.gov/energysaver/articles/tips-appliances → pages 94[226] M. P. Mills, “The cloud begins with coal,” Aug. 2013. [Online]. Available:http://www.tech-pundit.com/wp-content/uploads/2013/07/Cloud Begins With Coal.pdf?c761ac→ pages 94, 96126[227] J. Koomey, “Does your iPhone use as much electricity as a new refrigerator? Not even close.”Aug. 2013. [Online]. Available:http://thinkprogress.org/climate/2013/08/25/2518361/iphone-electricity-refrigerator/ → pages94[228] H. I. Bensch, S. Pigg, K. Koski, and R. Belshe, Electricity Savings Opportunities for HomeElectronics and Other Plug-In Devices in Minnesota Homes: A technical and behavioral fieldassessment. Energy Center of Wisconsin, 2010. → pages 96, 168[229] B. Urban, V. Tiefenbeck, and K. Roth, “Televisions, computers, and set-top boxes: The big threeof 2010 home consumer electronics energy consumption,” Fraunhofer Center for SustainableEnergy Systems, Cambridge, MA, 2011. → pages 96[230] B. Chrisman, H. Singh, and K. Rider, “2012 Appliance Efficiency Regulations,” CaliforniaEnergy Commission, California, Oct. 2012, CEC-400-2012-019. [Online]. Available:http://www.energy.ca.gov/2012publications/CEC-400-2012-019/CEC-400-2012-019-CMF.pdf→ pages 107[231] C. Wilson and H. Dowlatabadi, “Models of decision making and residential energy use,” AnnualReview of Environment and Resources, vol. 32, no. 1, pp. 169–203, 2007. → pages 107[232] T. Cooper, S. Fallender, J. Pafumi, J. Dettling, S. Humbert, and L. Lessard, “A semiconductorcompany’s examination of its water footprint approach,” in 2011 IEEE International Symposiumon Sustainable Systems and Technology (ISSST), May 2011, pp. 1–6. → pages 108[233] The Green Grid, “Water usage effectiveness (WUETM): a green grid data center sustainabilitymetric,” 2011, White Paper #35. [Online]. Available:http://www.thegreengrid.org/∼/media/WhitePapers/WUE → pages 108[234] C. E. H. Beaudrie and M. Kandlikar, “Horses for courses: risk information and decision makingin the regulation of nanomaterials,” Journal of Nanoparticle Research, vol. 13, no. 4, pp.1477–1488, Apr. 2011. → pages 108[235] Global e-Sustainability Initiative and The Boston Consulting Group, Inc, “GeSI SMARTer2020: The role of ICT in driving a sustainable future,” Dec. 2012. [Online]. Available:http://gesi.org/SMARTer2020 → pages 108[236] ITRS, “International technology roadmap for semiconductors, 2009 edition, executivesummary.” ITRS, 2009. → pages 135[237] M. L. Socolof, J. G. Overly, and J. R. Geibig, “Environmental life-cycle impacts of CRT andLCD desktop computer displays,” Journal of Cleaner Production, vol. 13, no. 1314, pp.1281–1294, Nov. 2005. → pages 142[238] M. LaPedus, “Industry agrees on first 450-mm wafer standard,” EETimes, Oct. 2008. [Online].Available: http://www.eetimes.com/electronics-news/4079657/Industry-agrees-on-first-450-mm-wafer-standard → pages 145[239] P. Apte, W. Bottoms, W. Chen, and G. Scalise, “Good things in small packages,” Spectrum,IEEE, vol. 48, no. 3, pp. 44 –49, Mar. 2011. → pages 147, 148[240] iFixit, “Apple A4 teardown.” [Online]. Available:http://www.ifixit.com/Teardown/Apple-A4-Teardown/2204/1 → pages 147127[241] I. Feldman, “iPad memories.” [Online]. Available: http://hightechbizdev.com/2011/03/24/ira-feldman-high-technology-business-development-ipad-memories/ → pages 148[242] K. Marsal, “Teardown of Apple’s 4th-gen iPod touch finds 256MB of RAM,” Sep. 2010.[Online]. Available: http://www.appleinsider.com/articles/10/09/08/teardown of apples 4th gen ipod touch finds 256mb of ram.html → pages 148[243] EveryiPod.com, “iPod Q&A – Updated October 11, 2011,” Oct. 2011. [Online]. Available:http://www.everymac.com/systems/apple/ipod/ipod-faq/differences-between-4th-gen-ipod-touch-3rd-gen-ipod-touch.html → pages 148[244] Chipworks Inc., “Teardown of the 3rd generation Amazon Kindle DX,” Nov. 2010. [Online].Available: http://www.chipworks.com/en/technical-competitive-analysis/resources/recent-teardowns/2010/11/teardown-of-the-3rd-generation-amazon-kindle-dx/ → pages 148[245] N. Horowitz, “New PS4 and Xbox One game consoles: A mixed bag for energy efficiency,”Natural Resources Defense Council Staff Blog, 2013. [Online]. Available:http://switchboard.nrdc.org/blogs/nhorowitz/new ps4 and xbox one game cons.html → pages156[246] J. Malmodin, D. Lunden, M. Nilsson, and G. Andersson, “LCA of data transmission and IP corenetworks,” in Electronics Goes Green 2012+ (EGG), 2012, 2012, pp. 1–6. → pages 158[247] D. Schien, C. Preist, M. Yearworth, and P. Shabajee, “Impact of location on the energy footprintof digital media,” in 2012 IEEE International Symposium on Sustainable Systems andTechnology (ISSST), May 2012, pp. 1–6. → pages 158[248] J. Baliga, R. Ayre, K. Hinton, W. V. Sorin, and R. S. Tucker, “Energy consumption in optical IPnetworks,” Journal of Lightwave Technology, vol. 27, no. 13, pp. 2391–2403, Jul. 2009. →pages 158[249] A. Fehske, G. Fettweis, J. Malmodin, and G. Biczok, “The global footprint of mobilecommunications: The ecological and economic perspective,” IEEE Communications Magazine,vol. 49, no. 8, pp. 55 –62, Aug. 2011. → pages 158, 159, 160[250] GSM Association, “Mobiles green manifesto 2012,” London, UK, Jun. 2012. [Online].Available: http://www.gsma.com/publicpolicy/mobilesgreenmanifesto → pages 159[251] J. Malmodin, D. Lunde´n, A. Moberg, G. Andersson, and M. Nilsson, “Life cycle assessment ofICT networks carbon footprint and operational electricity use from the operator, national andsubscriber perspective,” Journal of Industrial Ecology, 2014, in press. → pages 159, 161, 162[252] G. Auer, O. Blume, V. Giannini, I. Godor, M. Imran, Y. Jading, E. Katranaras, M. Olsson,D. Sabella, and P. Skillermark, “D2. 3: Energy efficiency analysis of the reference systems,areas of improvements and target breakdown,” EARTH project, 2010. [Online]. Available:https://www.ict-earth.eu/ → pages 160[253] DatacenterDynamics, “DatacenterDynamics global industry census 2012,” London, 2012.[Online]. Available: http://dcdi.datacenterdynamics.com/intelligence-report-2012 → pages 162[254] Cisco Systems Inc., “VNI overall forecast highlights,2012-2017,” Cisco Systems Inc., Report,2013. → pages 163128[255] Netflix, Inc., “Manage bandwidth usage,” netflix Help Center. [Online]. Available:https://help.netflix.com/en/node/87 → pages 165[256] GFK, “MultiMedia mentor,” 2012. [Online]. Available: http://www.gfk.com/us/news-and-events/News/Pages/Social-Media-Dominates-Smartphone-Internet-Time.aspx →pages 167[257] GSMA Intelligence, “Smartphone users spending more ’face time’ on apps than voice calls orweb browsing,” 2013. [Online]. Available: https://gsmaintelligence.com/analysis/2011/03/smartphone-users-spending-more-face-time-on-apps-than-voice-calls-or-web-browsing/271/→ pages 167[258] Temkin Group, “Data snapshot: Media use benchmark, 2013,” Mar. 2013. [Online]. Available:http://www.temkingroup.com/research-reports/data-snapshot-media-use-benchmark-2013/ →pages 167[259] T. Beauvisage, “Computer usage in daily life,” in Proceedings of the 27th internationalconference on Human factors in computing systems. Boston, MA, USA: ACM, 2009, pp.575–584. → pages 167[260] Ofcom, “The communications market 2013,” Aug. 2013. [Online]. Available:www.ofcom.org.uk/cmr → pages 167129Appendix ASupplementary material for Chapter 3A.1 Summary data tablesThis section contains data tables for each figure in the text.Percentage share by life cycle phaseManufacture Distribution Operation End-of-lifePrimary energy Total [MJ]Williams 2004 84.3 5.9 9.8 0 5924Atlantic Consulting 1998 28.8 0.1 70.4 -0.7 5334Braune and Held 2006 [EPIC-ICT] 19.2 0 78.8 -1.9 –Hischier et al 2007 [ecoinvent] 28.9 0.3 70.7 0.1 15011IVF 2007 13.9 2.3 83.6 -0.2 16165Masanet et al 2005 57.6 0 42.3 0.1 10113Kenma 2005 [MEEUP] 13.7 5.3 80.5 0.5 20876Global warming potential Total [kg CO2e]Atlantic Consulting 1998 32.7 0.1 66.3 -0.9 249Braune and Held 2006 [EPIC-ICT] 30.2 0 67 -2.8 –Hischier et al 2007 [ecoinvent] 33.7 0 65.6 0.7 775Choi 2006 89 1 8 2 –Duan 2009 29.4 0 63.7 -6.9 –IVF 2007 18.1 3.7 78.1 -0.1 761Tekawa 1998 (includes CRT) 18.1 0.7 80.9 0.3 717Apple 2010b (Mac Pro) 42 3 54 1 1880Apple 2010a (Mac Mini) 57 3 39 1 270Masanet et al 2005 57.6 0 42.3 0.1 1111Hikwama 2005 34.1 0 63.7 2.2 –Table A.1: Data table for Figure 3.1: Total impacts by life cycle phase130Williams2004[MJ]BrauneandHeld2006[MJ]IVF2007[MJ]Hischieretal2007[MJ]Duan2009[EcoIndicatorpts.]Tekawa1998[kgCO2e]Hikwama2005[kgCO2e]Mainboard: ICs 37.5 0 24.2 26.7 0 0 25.2Mainboard: other 9.5 53 44.5 38.9 55.7 73.2 69.8Power supply 0 13 0 11.7 11 4.9 0Drives 9.9 8 0 10.9 19.9 3.7 0Casing 7.3 13 16.2 5.8 8.5 8.5 1.8Other 35.8 13 15.1 6.1 5 9.8 3.3Table A.2: Data table for Figure 3.2: Manufacturing impacts by subassemblyWilliams(2004)IVF(2007)Hischieretal.(2007)Hikwama(2005)Component mass per desktop [kg]Total desktop 9 10.5 11.3 10Casing 6.4 7 6.1 5.5Mainboard 2 1 1.4 1Drives 0.6 – 1.5 2Power supply – – 1.7 1.1GWP per component [kg CO2e/kg]Total desktop – 13.2 22.7 –Casing – 3.2 2.4 –Mainboard – 63.8 71.6 –Drives – – 18.7 –Power supply – – 17.6 –Primary energy per component [MJ/kg]Total desktop 587 183 390 –Casing 61 44 42 –Mainboard 252 883 1226 –Drives 935 – 320 –Power supply – – 302 –Table A.3: Data table for Figure 3.3: Mass and impact factors by subassembly131Graph Study UEC Data Home or Data typeindex [kWh/yr] quality office1 Roth et al. (2002) (high-power PC) 717 Medium Office Secondary2 Braune and Held(2006) [EPIC-ICT] (gamer PC) 587 Low Home Primary3 US EPA(2010) [Energy Star, baseline] 408 Medium Both Secondary4 US EPA(2005) [Energy Star, baseline] 354 Medium Office Secondary5 Braune and Held (2006) [EPIC-ICT] 331 Low Office Primary6 Kemna (2005) [MEEUP] 322 High Home Mixture7 Moorefield et al. (2008) 305 High Office Primary8 Roth et al. (2002) 298 Medium Office Secondary9 US EPA (2005) [Energy Star, baseline] 285 Medium Home Secondary10 Kemna (2005) [MEEUP] 281 High Office Mixture11 Braune and Held (2006) [EPIC-ICT] 279 Low Office Primary12 US EPA(2010) [Energy Star, certified] 275 Medium Both Secondary13 US EPA(2005) [Energy Star, certified] 271 Medium Office Secondary14 Porter et al. (2006) 256 High Home Primary15 Hischier et al. (2007) [EcoInvent] 238 Low Office Mixture16 Roth and McKenney(2007) 236 Medium Home Secondary17 Duan et al. (2009) 223 Medium Both Secondary18 US EPA(2005) [Energy Star, certified] 214 Medium Home Secondary19 Kawamoto et al. (2001) 213 Low Office Primary20 Hischier et al. (2007) [EcoInvent] 204 Low Home Mixture21 MTP(2006) 196 High Home Primary22 IVF(2007) 194 High Office Mixture23 IVF(2007) 141 High Home Mixture24 Schlomann et al. (2005) 122 Medium Office Mixture25 Atlantic Consulting(1998) 110 Low Office Secondary26 Nordman and Meier(2004) 93 Low Home Mixture27 Schlomann et al. (2005) 78.1 Medium Home Mixture28 Choi et al. (2006) 76.1 Medium Office Secondary29 Hikwama(2005) 75.1 Low Home Secondary30 Williams(2004) 53.7 Low Office Secondary31 Choi et al. (2006) 49.1 Medium Home Secondary32 Kawamoto et al. (2001) 49 Low Home PrimaryTable A.4: Data table for Figure 3.4: Desktop unit energy132Graph Study Lifespan Includes Data Data typeindex (years) re-use? quality1 JEITA(2006) (Home) 8.4 Includes reuse Medium Primary2 Babbitt et al. (2009) (1990 purchases) 7.86 First life only High Primary3 Tekawa et al. (1997) (Office) 7 Not specified Low Secondary4 IVF(2007) 6.6 Includes reuse Medium Primary5 Oguchi et al. (2006) 6.6 Includes reuse Medium Primary6 Duan et al. (2009) 6 Includes reuse Low Secondary7 Kemna(2005) 6 Includes reuse Medium Primary8 Babbitt et al. (2009) (1995 purchases) 5.88 First life only High Primary9 JEITA(2006) (Office) 5.8 Includes reuse Medium Primary10 Babbitt et al. (2009) (2000 purchases) 5.49 First life only High Primary11 Tekawa et al. (1997) (Home) 5 Not specified Low Secondary12 Hikwama(2005) 5 First life only Low Secondary13 ESRI(2007) 4.3 Not specified Medium Primary14 Braune and Held(2006) [EPIC-ICT] 4 Not specified Medium Secondary15 Hischier et al. (2007) [EcoInvent] 4 First life only Medium Secondary16 Choi et al. (2006) 4 Includes reuse Medium Secondary17 Masanet et al. (2005) 4 Not specified Low Secondary18 Williams(2004) 3 First life only High Secondary19 Atlantic Consulting(1998) 3 First life only Low Secondary20 Smulders(2001) 3 First life only High Primary21 Williams and Hatanaka(2005) 2.99 First life only High PrimaryTable A.5: Data table for Figure 3.5: Desktop lifespanPrimary energy[MJ/kWh]Global warmingpotential[kg CO2e/kWh]LCA studiesWilliams 2004 3.6 –Atlantic Consulting 1998 11.5 0.51Hischier et al 2007 12 0.58IVF 2007 10.8 0.47Kemna 2005 10.5 0.46Masanet et al 2005 – 0.4Average (excl. Williams) 11.2 0.47Countries (ecoinvent data)Switzerland 10.6 0.13USA 13.9 0.83Europe-UCTE 12.8 0.59Germany 12.6 0.72France 13.6 0.11UK 12.4 0.68Norway 5.2 0.05China 14.1 1.45Japan 13.1 0.6Table A.6: Data table for Figure 3.6: Electricity impact factors133A.2 Mapping to subassembly categoriesThe list below briefly discusses interpretation of LCA studies when expressing their results in terms ofsubassemblies in Section 3.4.1.• Tekawa et al [160]: Component-level information is available in terms of total share of manufac-turing impacts only; the mapping to this study’s categories is straightforward.• Williams [13]: The study includes a substantial amount of “other” impacts due its focus on up-stream manufacturing steps. In grouping these results into component categories, “integrated cir-cuits” included Williams’ estimates for semiconductors, silicon wafers, electronic chemicals, andsemiconductor manufacturing equipment. The “mainboard:other” category consisted of “printedcircuit boards”, “passive components”, and a subset of “bulk materials–control unit”. The “drives”category consisted of “disk drives and other parts” and a subset of “bulk materials–control unit”.The casing consisted of a subset of “bulk materials–control unit”. The bulk materials are dividedaccording to Williams’ descriptions in the paper’s supporting information. Where materials areshared, such as aluminum, which is split between hard drives and circuit boards, the total impactsare divided equally among the identified categories. This study’s categorization of these resultsis imperfect; in particular, mainboard impacts are significantly lower than other studies. It ispossible some of the impacts apportioned to semiconductors should be applied to the mainboardinstead.• Hikwama [165]: Components in the inventory are grouped by sub-assembly with a straightfor-ward mapping to the component categories in this study.• Braune and Held [168]: A breakdown of manufacturing impacts by subassembly type with astraightforward mapping to categories used in this study, but in relative terms only; absolutenumbers were not provided.• Hischier et al [95]: The data are organized into sub-processes which map very closely to thecomponent categories used in this study.• IVF [169]: The product inventory is not characterized hierarchically, making it impossible toassign impacts to sub-assemblies like disk drives and power supplies. To map these impacts to134component categories in this study, it was assumed that sheet steel and ABS plastic belong to thehousing; all surface mount devices, printed wiring boards, solder, and “big capacitors and coils”belong to the mainboard; and all remaining materials (except integrated circuits, which can bemapped directly) are grouped as “other”.• Duan et al [159]: A breakdown of manufacturing impacts by component type is provided inrelative terms only, with a straightforward mapping to this study’s categories.A.3 Integrated circuit contentLCA studies represent integrated circuit content in three different units: area of silicon wafer input;area of finished die output, and mass of packaged chip. The text converts all identified inventories intofinished die area; this section details the calculations behind this conversion. Note the energy and otherimpacts in the chip package (and the packaging process) are assumed to be negligible compared to thoseof the silicon die.Input silicon wafer is converted to finished die through processing, but some area is wasted due todefective wafers and defective dies. Boyd et al. [172], summarizing industry data for a 45nm process,reports a typical yield of 88% for wafers (finished wafers/started wafers), and a yield of 443 good chipsper wafer with average size of 111 mm2; total functional area is thus 49,000 mm2 per wafer. As thisprocess uses 300 mm wafers with a total area of about 70,000 mm2, and 12% of wafers are discarded,this means one unit of wafer input produces about 0.6 units of finished die. Comparable estimates of0.55 [95] and 0.63 [12] are found in the literature. Thus, Williams [13] estimate of 110 cm2 wafer inputper desktop is equivalent to about 66 cm2 of finished die per desktop.Several studies measure IC content by mass of packaged ICs, and include an estimate of the silicondie mass as a percentage of the packaged IC mass, ranging from 0.9% to 5%. The area of a silicondie per unit mass can be determined by dividing the the standard thickness of a silicon wafer, 0.078cm [236], by the mass density of silicon, 2.32 g/cm3, which gives the area of a finished die to be about5 cm2/g.Supporting information for the ecoinvent database [95] reports silicon die area per unit mass to be2.7 cm2/g for logic chips (2% Si die by mass) and 1.2 cm2/g for memory chips (0.9% Si die by mass);these differ from the expected 5 cm2/g which should apply to all silicon die of thickness 0.078 cm.135Consequently, when the ecoinvent database attempts to supply a quantity of silicon die using a massratio, it uses an erroneously low estimate for die area. The source of the discrepancy here is an apparentconfusion between finished die area and package area in the calculations. See Section B.4 in AppendixB for further discussion.Only one study, Yao et al. [20], provides measurements directly in terms of finished die; these wereobtained using laboratory x-ray measurements of a mainboard with packaged chips. This study is largelya response to Williams [13] and attempts to correct Williams’ estimates by calculating primary energyconsumption using the same methods, but with a different estimate for semiconductor content. Theyappear to make two errors, however; first, they apply Williams’ estimate of energy per area of inputwafer to their measurements of area per finished die, neglecting to account for the 0.6 conversion factordiscussed above; second, they report Williams’ estimate of semiconductor energy requirements to be909 MJ per desktop, but neglect to add the additional terms specified by Williams for silicon wafers,electronic chemicals, and semiconductor manufacturing equipment, which bring the total to 1992 MJ.Both of these errors make Yao’s reported estimate of 99 MJ per desktop or 8.25 MJ per cm2 of finisheddie lower than they should otherwise be.136Appendix BSupplementary material for Chapter 4B.1 Summary data tablesThis section contains data tables summarizing all results for each graph in the text.Category Unit Desktop(ei)Desktop-towerDesktop-smallThinclientLCDmonitor,17”(ei)LCDmonitor,21.5”Laptop,withdock,12”(ei)Laptop,16”Netbook,10”iPadiPodtouchKindleRackserverSwitchPower supply Mass [g] 1463 1461 476 182 114 173 537 433 297 89 89 89 2911 193Casing Mass [g] 6207 6171 1258 860 862 2157 1413 904 265 151 49 59 8767 1404Circuit boards Mass [g] 1558 1028 407 227 94 38 345 281 127 30 6 33 2199 460IC’s (packages) Mass [g] 119 40 35 8 19 2 86 21 17 2 1 2 88 47IC’s (die) Area [mm2] 2195 500 218 126 355 22 1577 483 463 170 52 66 1683 366IC’s (die) Mass [mg] 3970 905 394 228 642 39 2851 873 837 307 94 120 3043 663Battery Mass [g] 0 0 0 0 0 0 273 244 178 129 16 51 0 0Display Mass [g] 0 0 0 0 4006 2350 328 553 204 342 35 34 0 0Other Mass [g] 1796 1959 796 0 0 350 367 334 250 34 3 45 1506 16Total Mass [g] 11144 10660 2972 1277 5096 5070 3349 2770 1337 777 198 312 15471 2119IC: integrated circuit. Circuit boards category excludes IC’s.Table B.1: Summary data table, product composition by mass137Category Desktop(ei)Desktop-towerDesktop-smallThinclientLCDmonitor,17”(ei)LCDmonitor,21.5”Laptop,withdock,12”(ei)Laptop,16”Netbook,10”iPadiPodtouchKindleRackserverSwitchPower supply 38.2 40.5 17.2 5.1 10.9 10.1 2.9 9 7.8 2.3 2.3 2.3 89.3 10.8Casing 13.1 15.3 2.9 3.7 6.8 10.2 64.4 6.8 2 1.3 0.4 0.5 18.5 3.8Circuit boards 76.9 38.6 23.7 12.6 9.6 2.2 32 16.2 8.6 1.7 0.4 2.4 128.8 25.3IC’s (packages) 80.4 15 11.2 5.6 14.2 1.2 58.9 15.5 12.1 1.3 0.6 1.2 50.8 34.3IC’s (die) 79 21.5 9.4 5.4 17 1 67.8 20.8 19.9 7.3 2.2 3.6 72.4 15.8Battery 0 0 0 0 0 0 1.4 1.3 0.9 0.7 0.1 0.3 0 0Display 0 0 0 0 236.3 138.6 19.3 31.9 6.7 9.9 0.8 2 0 0Other 29.9 28.9 7.2 0 0 1.9 7.4 4.9 2.9 0.1 0 0.3 18.2 0.2Transport 2.8 2.7 0.8 0.3 1.3 1.3 0.8 0.7 0.3 0.2 0 0.1 3.9 0.5Assembly 1.2 1.2 1.2 0.9 1.2 1.2 1.2 0.9 0.9 0.6 0.6 0.6 1.2 1.2Total 321.6 163.7 73.5 33.6 297.3 167.8 256.2 107.8 62.2 25.5 7.5 13.3 383.1 91.8IC: integrated circuit. Circuit boards category excludes IC’s.Table B.2: Summary data table, embodied GHG emissions, in kg CO2eCategory Desktop(ei)Desktop-towerDesktop-smallThinclientLCDmonitor,17”(ei)LCDmonitor,21.5”Laptop,withdock,12”(ei)Laptop,16”Netbook,10”iPadiPodtouchKindleRackserverSwitchPower supply 723 754 314 96 201 190 54 173 151 42 42 42 1653 203Casing 215 273 49 70 121 203 330 134 39 22 7 8 302 58Circuit boards 1621 757 439 233 180 41 606 323 161 32 7 45 2407 480IC’s (packages) 1778 301 229 126 269 24 1259 306 245 25 11 22 1008 650IC’s (die) 1223 318 138 80 256 16 1003 307 294 108 33 53 961 233Battery 0 0 0 0 0 0 28 25 18 13 2 5 0 0Display 0 0 0 0 3518 2064 288 475 101 148 11 30 0 0Other 536 516 132 0 0 41 133 88 54 3 0 6 342 4Transport 45 43 12 5 20 20 13 11 5 3 1 1 62 9Assembly 25 25 25 19 25 25 25 19 19 13 13 13 25 25Total 6165 2987 1338 630 4590 2624 3738 1861 1088 410 127 225 6761 1661IC: integrated circuit. Circuit boards category excludes IC’s.Table B.3: Summary data table, cumulative energy demand, in MJ138Product (this study) Mass [g] GHG [kg CO2e] Primary energy [MJ]Desktop (ei) 11144 321.6 6165.2Desktop – tower 10660 163.7 2986.8Desktop – small 2972 73.5 1338.3Thin client 1277 33.6 630.1LCD monitor, 17” (ei) 5096 297.3 4589.5LCD monitor, 21.5” 5070 167.8 2624.1Laptop, with dock, 12” (ei) 3350 256.2 3738.2Laptop, 16” 2770 107.8 1861.1Netbook, 10” 1337 62.2 1087.8iPad 777 25.5 409.6iPod touch 198 7.5 127.1Kindle 312 13.3 225.1Rack server 15471 383.1 6761.1Switch 2119 91.8 1660.8Table B.4: Summary data table, mass, embodied GHG, and embodied primary energy demand,this studyApple Product Mass [g] GHG [kg CO2e]27-inch LED Cinema Display 10900 516.6Thunderbolt Display 10922 301.6Apple TV 270 25.221.5-inch iMac 9304 349.227-inch iMac 13800 455.7iPad 2 590 63iPhone 3GS 135 24.75iPhone 4 135 25.65iPod classic 139 11.5iPod nano 21.1 7.02iPod shuffle 12.5 2.88iPod touch 101 15.3711-inch MacBook Air 1080 16213-inch MacBook Air 1350 198.4MacBook 2132 142.613-inch MacBook Pro 2041 20315-inch MacBook Pro 2539 289.817-inch MacBook Pro 2994 334.8Mac mini with Lion Server 1400 146.9Mac mini 1300 153.9Mac pro 18100 789.6Xserve 13540 416Table B.5: Summary data table, mass and embodied GHG, Apple dataset139B.2 Adjustments and comparison to ecoinventThis study used the data and assumptions from the ecoinvent database, with a few adjustments andmodifications. Three products from the ecoinvent database [191], a desktop (‘desktop computer, withoutscreen, at plant’), laptop (‘laptop computer, at plant’), and LCD monitor (‘LCD flat screen, 17 inches, atplant’), were re-implemented in our modeling framework. This study’s results can be compared againstthe original results from the ecoinvent database both to illustrate the effects of the adjustments, and toconfirm that this study’s framework reproduces the ecoinvent modeling assumptions. A summary ofthe original results and our adjusted results is in Table B.6 below, followed by a summary of the majoradjustments made in this study.Parts were categorized as follows: circuit boards includes mainboards, RAM, video cards, and anyother circuit boards, as well as all integrated circuits, connectors, capacitors, processor heat sinks, andother board-mounted components; casing includes all metal and plastic frames and screws from thedevice exterior; display is the screen unit only; power supply includes power cables as well as internalpower supplies and smaller external supplies including chargers; battery includes only large cells, suchas lithium-ion laptop batteries; and other includes any remaining components such as interior powercables, internal frames, disk and optical drives, case fans, and any remaining parts.Desktop Laptop LCD monitorOriginal Adjusted Original Adjusted Original AdjustedAssembly 1.4 1.2 0.9 1.2 52.7 1.2Battery 1.6 1.4Casing 15.6 13.1 61 64.4 6.9 6.8Circuit boards 180.1 236.4 104 158.7 34.3 40.8Display 19.3 19.3 236.3 236.3End of life 5.5 3.9Other 32.7 29.9 9.1 7.4Packaging 2.8 2.9 3.1Power supply 28.4 38.2 3.8 2.9 10.9Transport 2.7 2.8 0.8 0.7 1.3Total 269.3 321.6 207.4 256 333.3 297.3Table B.6: Global warming potential results for original ecoinvent study and this study’s adjust-ments, in kg CO2e140B.2.1 System boundaryThis study excludes the use and end-of-life phases in order to focus on impacts during device production.Device packaging is excluded due to lack of data for packing for some of the products being studied;ideally packaging would be included in the study boundary for all products, but the magnitude of impactsdue to packaging are typically small, so its exclusion should not significantly affect the results. Finalassembly and transport are included using the standard ecoinvent assumptions. Note that the LCD screenin ecoinvent did not include transport, and included an extra assembly step which I believe is spurious(see discussion below); I have adjusted this study so it is consistent with the others.B.2.2 Silicon die and integrated circuitsSilicon die content is calculated using an empirically derived relationship, discussed in Section B.4. Thesilicon die content per packaged chip is estimated to be about 18 mm2 of die per gram of packaged chip,which is significantly larger than the ecoinvent model, which estimates 5.5 mm2 per gram for logic chips(“integrated circuit, IC, logic type, at plant”) and 10.1 mm2 die per gram for memory chips (“integratedcircuit, IC, memory type, at plant”). The lower estimates in ecoinvent are due to apparently erroneousassumptions in the ecoinvent database, discussed in previous work [1, 133] and in Section B.4. Thedesktop, laptop, and LCD monitor models have been adjusted to use the silicon die models in this study,which has significantly increased the impacts due to integrated circuits. In addition, updated life cycleassessment results for silicon die have been applied [172], though this has a relatively small influenceon the results.B.2.3 LCD assemblyThe ecoinvent process “LCD flat screen, 17 inches, at plant” includes an assembly process called “as-sembly, LCD screen” which includes significant chemical usage. However, the top-level process forthe LCD monitor also includes an LCD module (“LCD module, at plant”) which itself contains a verysimilar assembly process (“assembly, LCD module”). The two processes, “assembly, LCD screen” and“assembly, LCD module”, are in fact nearly identical in their contents, except the inventory contentsin the latter are 3.91 times larger in all cases, and there are a small number of items present in the lat-ter process but not in the former. The ecoinvent documentation is not fully clear on the purpose andfunction of the “assembly, LCD screen” process.141There is no reason to suppose that additional chemical-intensive processes are required in orderto assemble the completed components of an LCD monitor (casing, LCD module, cables, etc.); thefinished product can be disassembled into such components using a hand screwdriver. Therefore, eitherthe chemical-intensive “assembly, LCD screen” process is spurious, or the “LCD module, at plant”process is intended to represent only part of the central display apparatus. I believe the former is morelikely, for two reasons. First, this would be more consistent with the conventions of ecoinvent in whichtop-level processes include finished components; notably, the laptop computer process includes thisLCD module, but not the “assembly, LCD screen” process. Second, the study upon which this datawas originally drawn defines an “LCD module” as including “the LCD panel (i.e., front and back glasspanels, liquid crystals and polarizers, column and row drivers), the backlight unit, and the main LCDcontroller PWB” [237]. This represents all of the major components in an LCD monitor [237] whichimplies that the LCD module is indeed the entire central display apparatus.In order to maintain a consistent approach I have assumed that the “LCD module, at plant” processrepresents a finished display component, and that products which include this component do not requirean additional chemical-intensive assembly step. Therefore, I have adjusted the “LCD flat screen, 17inches, at plant” process to remove the “assembly, LCD screen” component. This process accounted for53.0 kg CO2e of the LCD flat screen’s global warming potential.B.2.4 LCD power supplyThere is no power supply modeled in the ecoinvent LCD monitor process. However, the main cir-cuit board process is defined as a mixture of a surface-mount populated printed wiring board and athrough-hole populated printed wiring board. The surface-mount board modeled in ecoinvent containscomponents which suggest it is used for logic applications, while the through-hole board contains powerelectronics which suggest it is used in power supplies. In order to maintain consistency with other prod-ucts, I have split these two components in my adjusted study and assumed that the through-hole circuitboard process models the monitor’s power supply.B.2.5 Other variationThe above adjustments account for nearly all of the differences between this study’s results and theoriginal ecoinvent studies. The remainder is caused by this study’s modeling framework which has142condensed the ecoinvent database into a subset of about 100 important processes. In some cases there isa reduction of detail in which I have chosen to use proxy materials; for example, I model only one typeof radial cylindrical inductor, whereas ecoinvent has separate processes for small and large inductors.The reduction of detail simplifies this project with minimal loss of accuracy.B.3 Uncertainty factorsThe ecoinvent database uses a semi-quantitative scale based on pedigree matrices to assign uncertaintydistributions[191]. In general, if a product includes n kg of a substance, then n is assumed to be themean of a random variable N with standard deviation s, where s depends on the quality of data whichproduced the estimate of n kg, and the value of s is a function of expert judgments and is determinedusing a pedigree matrix.Operationally, the ecoinvent documentation describes several uncertainty factors, Ui, which quanti-tatievly score modeling uncertainty in terms of reliability, completeness, temporal correlation, geograph-ical correlation, further technological correlation, and sample size. When assessing the uncertainty ofa database model which is intended to model a physical component, the analyst refers to a pre-definedmatrix and scores the model in each of the above categories, such that higher-quality or less uncertainmodels receive lower scores. Each score is mapped to a numerical uncertainty factor using a predefinedscoring system. A quantity called the geometric standard deviation is defined, σ2g , which is obtainedthrough a numeric combination of uncertainty factors Ui. The geometric standard deviation is calculatedas σ2g = exp(U) where U =√∑i(ln2Ui). This standard deviation is applied along with the mean quan-tity to define a probability distribution for each quantity, which is generally assumed to be lognormalin the ecoinvent database. With such distributions applied throughout the model, it is possible to deriveoutput standard deviations through Monte Carlo analysis.This method has obvious limitations as it is based on expert judgments rather than empirical mea-surements and not does capture structural uncertainties in the model such as truncation error; this uncer-tainty analysis should therefore be understood as an attempt to bound that uncertainty which is internalto the ecoinvent modeling framework. Structural uncertainties that challenge the framework itself arenot considered.I have assigned standard deviations to all line items in each product’s bill of materials according tomy judgment of how well the ecoinvent processes, especially those modeling electronics components,143map to the components in the products under study. In the interests of simplicity, most quantities areassessed the same score, with the exception of silicon die and LCD screens. The scoring according tothe pedigree matrix format in the ecoinvent database is shown below in Table B.7. The scores are myassessment; the uncertainty factors are a function of the scores, determined by a mapping defined in theecoinvent documentation [191].Category Score Uncertaintyfactors UiReliability 3: Non-verified data partly based on qualified estimates 1.1Completeness 4: Representative data from only one site relevant for the market consid-ered OR some sites but from shorter periods1.1Temporal correlation 3: Less than 10 years of difference to our reference year (2000) 1.1Geographical corre-lation2: Average data from larger area in which the area under study is included 1.01Further technologicalcorrelation3: Data on related processes or materials but same technology, OR Datafrom processes and materials under study but from different technology1.2Sample size 5: unknown 1.2Table B.7: Pedigree matrix scoring for uncertainty characterization: assessment of ecoinvent com-ponent modelsUsing the above method and uncertainty factors in order to express the correlation of ecoinventelectronics component data as applied to the products in this study, a geometric standard deviationσ2g = 1.36 is obtained. This factor is applied as the default for all inventory items in this study. For silicondie and LCD screens which have additional uncertainty, the reliability and technological correlationscores are downgraded from 3 to 4, which changes the uncertainty factors from 1.10 and 1.20 to 1.20and 1.50, respectively, and results in a new geometric standard deviation σ2g = 1.65. Transport processesare assigned a score of σ2g = 2.1 in order to maintain consistency with other electronics processes inecoinvent.These standard deviations are sufficient to impose probability distributions on all components ineach product LCA model, with which Monte Carlo analysis was performed to produce the output stan-dard deviations described in the text. Numerical results of that analysis are shown in Table B.8.B.4 Silicon die content in integrated circuitsSilicon die are a key component contributing to emissions, but the precise quantity of silicon die in adevice can be difficult to estimate, even through teardown analysis, because the silicon die are usuallyencased in packages. Die can be identified through x-rays, removal of package through mechanical144Mean Standard deviation[kg CO2e] [kg CO2e] [% of mean]Rack server 378.6 40.9 0.11Desktop (ei) 307.9 31.6 0.1LCD monitor, 17” (ei) 295 45.4 0.15Laptop, with dock, 12” (ei) 249.3 44.8 0.18LCD monitor, 21.5” 167.4 31.8 0.19Desktop - tower 158 15.2 0.1Laptop, 16” 107.2 10.6 0.1Switch 91.4 8.8 0.1Desktop - small 72.7 7.6 0.1Netbook, 10” 60.6 7 0.12Thin client 34.9 3.4 0.1iPad 25.7 3.9 0.15Kindle 13.2 1.7 0.13iPod touch 7.5 0.9 0.12Table B.8: Monte Carlo analysis resultsgrinding (which is time-consuming, intricate work), or by approximation given the characteristics of thepackage. The ecoinvent database models follow the third approach based on a small set of empiricalmeasurements. In this study, a wider sample of packaged ICs were x-rayed in order to measure theirsilicon die content; this dataset, presented below, was used to develop higher-quality approximations ofsilicon die based on package characteristics which were applied in the study.Figure B.1 below shows the results of X-ray measurements of packaged ICs which were used toderive conversion ratios for estimating silicon die content. Twenty-two ICs were successfully X-rayedand their silicon die area measured. The dataset is split to illustrate that most surface-mount (SMT) ICsare very small and show high diversity in package types; the linear trend is mostly driven by large ball-grid array (BGA) ICs with large die. Errors are thus likely when estimating die content of small ICs, butthe absolute magnitude of these errors will also be small in proportion to the size of these ICs. Diversityin packaging technique leads to non-linear variation which is more pronounced in measurements ofmass; it is thus preferable to use the top-down area of the package to estimate die content.These results were used to create ratios that allow estimation of silicon die by packaged IC area andby IC mass. The mass of a die was calculated using the volume density of silicon, 2.33 g per cm3, andthe standard thickness of a 300 mm wafer, 775 µm [238]. Using a simple linear fit provides a best-fitmass ratio of about 18 mm2 of silicon die per gram of packaged IC (R2 = 0.58), and a best-fit area ratioof about 0.078 mm2 of silicon die per mm2 of packaged IC footprint area (R2 = 0.70). By comparison,145y = 18.429xR² = 0.57930204060801001201401600 2 4 6 8 10Silicon die area (mm^2)Packaged chip mass (g)large BGAsmdLinear (trend)y = 0.0778xR² = 0.69670204060801001201401600 500 1000 1500 2000Silicon die area (mm^2)Packaged chip area (mm^2)large BGAsmdLinear (trend)Figure B.1: Silicon die area measurements via X-ray, with regression linesthe ecoinvent database assumes 5.5 mm2 silicon die per gram of packaged IC for logic chips and 10.1mm2 silicon die per gram of packaged IC for memory chip.B.4.1 Discrepancy in ecoinvent modelsThe ecoinvent calculation uses a mass ratio to determine die size. Given an example packaged IC withdimensions of 27 mm * 27 mm * 2.36 mm, a mass of 2.62 g, and an assumption that the silicon die massis 2% of the packaged IC mass, the ecoinvent calculation of silicon die area per kilogram of packagedIC is as follows, from [95]:Die area per kg packaged IC = (0.027∗0.027)∗0.02/0.00262 (B.1)= 0.0056 [m2/kg] (B.2)The equation has the following factors:• Left side of equation: Die area / packaged IC mass• Right side of equation: Packaged IC area * (die mass / packaged IC mass) / (packaged chip mass)The units of each factor are as follows:• Left side of equation: m2 die/ kg packaged IC146• Right side of equation: (kg die/kg packaged IC) / kg packaged ICUnfortunately, the units on the right side of the equation do not resolve to those on the left side; thisis the source of the discrepancy. The difficulty could be resolved by replacing the mass ratio with a ratioexpressing die area divided by packaged chip area:• Left side of equation: m2 die / kg packaged IC• Right side of equation: m2 chip * (m2 die/m2 packaged IC) / kg packaged chipThe data above gives an area ratio of 0.078 mm2 die per mm2 packaged IC. Using the example chipwith area 27 mm * 27 mm and mass of 2.62 g would imply a die area of (0.027*0.027) * 0.078 / 0.00262= 0.021 m2 per kg of packaged IC, which is a factor of 3.9 higher than the ecoinvent calculated resultof 0.0056 m2 per kg packaged IC. Alternatively, the intended result, m2 die / kg packaged IC, can bedetermined empirically as I have done by examining the die area and packaged IC masses of severalpackaged ICs. The area ratio measured, 18 mm2 die per gram of packaged IC, or 0.018 m2 die per kg ofpackaged IC, is a factor of 3.2 higher than ecoinvent’s assumption of 0.0056 m2 die per kg of packagedIC. These results indicate that the ecoinvent dataset significantly underestimates the silicon die contentof packaged IC’s due to a calculation error. Thus, the area and mass ratios developed in this sectionare applied throughout the study instead of the ecoinvent models to approximate silicon die content inpackaged ICs.B.4.2 Stacked IC’sSome IC packages contain multiple die in order to save space and improve connectivity between mod-ules. The practice is especially common in mobile devices which require high performance but areheavily space-constrained; a common use is to combine multiple memory die, alone or with a processor,either by stacking silicon die directly on top of one another within one package, a technique known as“stacked die”, or by stacking packaged chips, a technique known as “package-on-package” [239]. Theassessment technique of x-raying packages was not able to clearly detect stacked die or package-on-package configurations, so I refer to industrial literature to identify these chips and make appropriatecorrections to our bills-of-materials.Third party analysis of the iPad showed that its A4 processor contains one CPU die and two memorydie [240]. Cross-sectional photographs from that analysis show that the RAM dies are 85% of the size147of the processor die in one dimension. Assuming the same is true in the other dimension, the RAM diearea is 72% of the processor die, which means that the approximately 50 mm2 CPU die is accompaniedby two RAM die with area about 36 mm2 each. The NAND solid-state storage chip did not includestacked die in the iPad model we analyzed, though higher-capacity iPad models would have [241]. The4th generation iPod touch uses the same A4 processor with memory die as the iPad [242]. The 3rdgeneration model we analyzed has a Samsung ARM processor with the same amount of RAM [243]and thus also likely uses a stacked configuration; the package is smaller, but we assume the same ratioof stacked memory die to processor die as in the iPad. The 3rd generation Kindle uses a low-cost ARMprocessor which does not appear to include stacked die or chips [244]. The use of stacked packagingtechnologies is unlikely in this product due to its low cost and lower performance requirements.Stacked-die packages are common in flash memory used in memory cards and in cell-phone pro-cessors [239]. Because of the added cost and complexity of using stacked die packaging, they are mostlikely to appear in small devices which are heavily area-constrained, and unlikely to appear in laptop anddesktop computers and other larger equipment. I found no third-party teardowns of the other productsin our study to compare against, but did search for information on each major IC (CPU and RAM) andfound no evidence to suggest that any stacked-die or package-on-package technologies were in use. Assuch, no further adjustments were made to account for stacking technologies.Technical and economic constraints point to future widespread usage of many types of stackingtechnologies, especially in mobile devices [239]. These trends will make identifying the die area usedin a product considerably more difficult. Mass of packaged chips would no longer be viable as a func-tional unit; life cycle assessments must accurately account for die area. X-raying would no longer be aviable technique for tabulating this area, unless very high-resolution cross-sectional x-rays were avail-able; grinding packages away to reveal the die inside will probably be the only reliable technique foraccurately measuring silicon die content of an integrated circuit. Ideally, manufacturer specificationscould be made available and incorporated into life cycle assessments in order to make such painstakingmeasurements unnecessary.B.5 Linear regression model selectionThree linear models are discussed in the text. Table B.9 are the fit results for the best fifteen modelswhich were attempted during the model selection process. The cross-validated sum of squares (cvss)148and raw and adjusted R2 are shown for each of two datasets; dataset one represents this study, and datasettwo is Apple’s dataset. The cvss scores are normalized to the mass-only model. The combined score isequal to the sum of the squares of the two cvss scores. Italicized rows represent the three models whichare presented in the text.Model predictors cvss1 cvss2 score R21 R21-adj. R22 R22-adj.Ma, int=0 1 1 2 0.85 0.84 0.89 0.88Bo, Ps, Di, Ba, int=0 0.38 1.39 2.09 0.97 0.96 0.96 0.96Bo, Di, Ba, int=0 0.42 1.4 2.12 0.97 0.96 0.94 0.93Bo, Di, Ba 0.45 1.4 2.17 0.92 0.9 0.87 0.84Ca 1.32 1.25 3.3 0.55 0.51 0.75 0.74Bo, Ps, Di, Ba, Ca, Ot 0.62 1.81 3.66 0.94 0.88 0.95 0.93Bo, Ps 1.36 1.91 5.5 0.55 0.47 0.68 0.64Ca, int=0 1.63 1.94 6.42 0.73 0.71 0.8 0.79Bo 1.24 2.36 7.11 0.55 0.51 0.59 0.57Bo, Vo 1.98 1.91 7.56 0.57 0.5 0.7 0.67Bo, Ps, Di, Ba, Ca, Ot, int=0 0.45 2.77 7.85 0.97 0.95 0.97 0.96Bo, Ps, Di 0.51 2.76 7.88 0.86 0.82 0.82 0.78Ca, Vo, int=0 2.3 1.94 9.04 0.74 0.69 0.82 0.8Bo, Ps, Di, Ba, Vo, int=0 0.56 3.01 9.37 0.97 0.95 0.96 0.95Bo, Ps, Di, Ba, Vo 1.69 2.77 10.49 0.93 0.89 0.93 0.91Ma: product mass; Bo: circuit board mass; Ps: power supply mass; Di: display massBa: battery mass; Ca: casing mass; Ot: other mass; Vo: product volume; int: y-intercept termModels discussed in the text are in italicsTable B.9: Model selection results for the top fifteen models.The circuit board + display + battery model, which had the third-best score, was selected because itsscore is very close to that of the model with the second-best score, and it requires one fewer predictor.The circuit board + power supply + display + battery + casing + other model was selected as an illustra-tive case because it had very low residual error (visible through its high R2) but performed poorly whencross-validated on Apple’s dataset. This suggests that the apparent additional precision supplied byadding more terms to that model is due to overfitting. Regression outputs for the three selected modelsare shown below in Tables B.10 and B.11.149Mass GHG Model estimates [kg CO2e][g] [kg CO2e] pcb+disp+batt Mass only All internalDesktop (ei) 11144 321.6 285.4 302.5 310.3Desktop – tower 10660 163.7 188.3 289.3 181Desktop – small 2972 73.5 74.5 80.7 64.9Thin client 1277 33.6 41.5 34.7 50.3LCD monitor 17-inch (ei) 5095 297.3 296.2 138.3 294.7LCD monitor 21.5-inch 5070 167.8 160.5 137.6 162.9Laptop with dock 12-inch (ei) 3349 256.2 167.6 90.9 171.4Laptop 16-inch 2770 107.8 161.8 75.2 162.2Netbook 10-inch 1337 62.2 90.7 36.3 83.9iPad 777 25.5 67.1 21.1 69.3iPod touch 198 7.5 8.2 5.4 2.8Kindle 312 13.3 23.8 8.5 21.5Rack server 15471 383.1 402.7 419.9 383.5Switch 2119 91.8 84.2 57.5 111.7Table B.10: Linear regression model outputs, this study’s datasetMass GHG Model estimates [kg CO2e][g] [kg CO2e] pcb+disp+batt Mass only All internal27-inch LED Cinema Display 10900 517 351.3 430.4 375.9Thunderbolt Display 10922 302 359.4 431.2 386.5Apple TV 270 25 16.2 10.7 0.321.5-inch iMac 9304 349 363.7 367.4 31727-inch iMac 13800 456 551.8 544.9 525.8iPad 2 590 63 74.2 23.3 68.9iPhone 3GS 135 25 18.1 5.3 17.2iPhone 4 135 26 17.2 5.3 16.8iPod classic 139 12 8.1 5.5 8.7iPod nano 21 7 1.9 0.8 1.1iPod shuffle 13 3 0.9 0.5 0iPod touch 101 15 10.6 4 7.111-inch MacBook Air 1080 162 128.3 42.6 11213-inch MacBook Air 1350 198 163.8 53.3 140.6MacBook 2132 143 204.8 84.2 193.713-inch MacBook Pro 2041 203 222.1 80.6 237.615-inch MacBook Pro 2539 290 264.3 100.2 272.117-inch MacBook Pro 2994 335 334 118.2 344.7Mac mini with Lion Server 1400 147 132.5 55.3 112.1Mac mini 1300 154 132.5 51.3 102Mac pro 18100 790 588.7 714.7 782.5Xserve 13540 416 588.7 534.6 428.1Table B.11: Linear regression model outputs, Apple’s dataset150Appendix CSupplementary material for Chapter 5C.1 Summary data tablesThis section provides numeric data tables corresponding to the graphs in the text, for ease of reference.Total emissions [Mt CO2e]2012 2017Mean stdev Mean stdevBy end-useOnline video 23.9 2.0 36.7 5.3Other online 31.4 2.2 35.1 4.5Offline 14.1 1.2 5.7 0.7IP video on demand 26.7 2.1 34.1 5.0Broadcast TV 74.8 7.9 65.4 13.5Total 170.9 8.8 177.0 16.0By locationBroadcast TV network 4.7 0.5 4.8 1.0Datacenter 27.4 3.2 33.6 7.1Fixed network 10.4 0.9 13.3 2.3Mobile network 6.2 0.8 5.1 1.1Devices 122.3 9.3 120.3 16.9Total 170.9 9.9 177.0 18.6Table C.1: Total US consumer emissions by end-use and location, with Monte Carlo results1512012 2017End-use Platform Location of emissions Energyperplatform[kWh/yr]GHGemb.perplatform[kgCO2e/yr]GHGtotalperplatform[kgCO2e/yr]GHGUSconsumertotal[MTCO2e/yr]Energyperplatform[kWh/yr]GHGemb.perplatform[kgCO2e/yr]GHGtotalperplatform[kgCO2e/yr]GHGUSconsumertotal[MTCO2e/yr]Broadcast TV TV platform Broadcast TV network 21.4 0.0 12.9 4.6 21.4 0.0 12.9 4.8Devices 285.6 26.9 198.2 71.0 247.8 22.8 171.5 63.7Total [TV, broadcast TV] 307.0 26.9 211.1 75.6 269.2 22.8 184.3 68.5IP VoD TV platform Datacenter 55.6 1.6 34.9 12.5 60.3 1.7 37.9 14.1Fixed network 20.8 1.0 13.5 4.8 24.1 1.4 15.8 5.9Devices 37.4 3.5 26.0 9.3 55.4 5.1 38.4 14.2Total [TV, IP VoD] 113.9 6.1 74.4 26.7 139.8 8.2 92.1 34.2Online video PC platform Datacenter 57.1 1.6 35.9 8.3 53.0 1.5 33.3 7.2Fixed network 21.2 1.0 13.7 3.2 20.8 1.2 13.7 3.0Mobile network 6.3 1.9 5.6 1.3 2.6 0.7 2.3 0.5Devices 29.9 5.3 23.2 5.4 39.8 6.9 30.8 6.7Total [PC, online video] 114.5 9.8 78.5 18.2 116.2 10.3 80.0 17.3Smartphone Datacenter 3.3 0.1 2.1 0.2 11.6 0.3 7.3 1.5Fixed network 0.8 0.0 0.5 0.1 3.2 0.2 2.1 0.4Mobile network 13.2 3.9 11.8 1.4 9.7 2.7 8.5 1.8Devices 0.1 0.6 0.6 0.1 0.4 1.6 1.8 0.4Total [Smartphone, online video] 17.4 4.6 15.0 1.8 24.9 4.8 19.7 4.1Tablet Datacenter 12.0 0.3 7.6 0.5 51.2 1.4 32.2 3.9Fixed network 4.3 0.2 2.8 0.2 19.2 1.1 12.6 1.5Mobile network 7.2 2.1 6.4 0.4 8.5 2.3 7.4 0.9Devices 2.0 10.2 11.4 0.7 6.1 28.1 31.7 3.8Total [Tablet, online video] 25.5 12.9 28.2 1.7 84.9 32.9 83.9 10.1TV platform Datacenter 4.6 0.1 2.9 1.0 9.0 0.3 5.7 2.1Fixed network 1.7 0.1 1.1 0.4 3.6 0.2 2.4 0.9Devices 3.4 0.3 2.4 0.8 9.8 0.9 6.8 2.5Total [TV, online video] 9.7 0.5 6.4 2.3 22.4 1.4 14.8 5.5Other online PC platform Datacenter 27.1 0.8 17.0 4.0 17.9 0.5 11.3 2.4Fixed network 10.0 0.5 6.4 1.5 6.9 0.4 4.6 1.0Mobile network 6.2 1.8 5.5 1.3 1.6 0.4 1.4 0.3Devices 101.0 17.8 78.4 18.2 115.9 20.1 89.6 19.4Total [PC, other online] 144.2 20.9 107.4 25.0 142.4 21.4 106.9 23.2Smartphone Datacenter 2.1 0.1 1.3 0.2 4.9 0.1 3.1 0.6Fixed network 0.4 0.0 0.2 0.0 1.1 0.1 0.7 0.1Mobile network 13.0 3.9 11.7 1.4 5.9 1.6 5.1 1.1Devices 1.7 7.8 8.8 1.1 2.3 10.1 11.5 2.4Total [Smartphone, other online] 17.2 11.8 22.1 2.7 14.1 12.0 20.4 4.21522012 2017End-use Platform Location of emissions Energyperplatform[kWh/yr]GHGemb.perplatform[kgCO2e/yr]GHGtotalperplatform[kgCO2e/yr]GHGUSconsumertotal[MTCO2e/yr]Energyperplatform[kWh/yr]GHGemb.perplatform[kgCO2e/yr]GHGtotalperplatform[kgCO2e/yr]GHGUSconsumertotal[MTCO2e/yr]Tablet Datacenter 6.0 0.2 3.7 0.2 17.9 0.5 11.3 1.4Fixed network 2.0 0.1 1.3 0.1 6.4 0.4 4.2 0.5Mobile network 7.1 2.1 6.3 0.4 5.1 1.4 4.5 0.5Devices 5.2 25.8 28.9 1.7 4.7 21.5 24.3 2.9Total [Tablet, other online] 20.2 28.2 40.3 2.4 34.1 23.7 44.2 5.3TV platform Datacenter 2.2 0.1 1.4 0.5 3.0 0.1 1.9 0.7Fixed network 0.8 0.0 0.5 0.2 1.2 0.1 0.8 0.3Devices 3.4 0.3 2.4 0.8 3.3 0.3 2.3 0.8Total [TV, other online] 6.4 0.4 4.2 1.5 7.5 0.5 4.9 1.8Offline PC platform Devices 56.1 9.9 43.6 10.1 17.3 3.0 13.4 2.9Smartphone Devices 1.2 5.6 6.3 0.8 0.3 1.3 1.5 0.3Tablet Devices 4.8 24.0 26.9 1.6 1.2 5.5 6.2 0.7TV platform Devices 6.8 0.6 4.7 1.7 6.5 0.6 4.5 1.7Total [Broadcast TV, all platforms] 75.6 68.5Total [IP Video on demand, all platforms] 26.7 34.2Total [Online video, all platforms] 24.0 37.0Total [Other online, all platforms] 31.6 34.5Total [Offline, all platforms] 14.2 5.6Total [All end-uses, all platforms] 172.1 179.8Table C.2: Energy and GHG emissions by end-use and platform1532012 2017Platform Location of emissions Energyperplatform[kWh/yr]GHGemb.perplatform[kgCO2e/yr]GHGtotalperplatform[kgCO2e/yr]GHGUSconsumertotal[MTCO2e/yr]Energyperplatform[kWh/yr]GHGemb.perplatform[kgCO2e/yr]GHGtotalperplatform[kgCO2e/yr]GHGUSconsumertotal[MTCO2e/yr]Desktop PC Datacenter 84.2 2.4 52.9 5.4 70.9 2.0 44.6 4.0Desktop PC Fixed network 31.5 1.5 20.4 2.1 28.3 1.6 18.6 1.7Desktop PC Devices 308.0 47.0 231.8 23.8 298.0 43.0 221.8 20.1Total 423.8 50.9 305.1 31.3 397.3 46.6 285.0 25.8Laptop (Wi-Fi) Datacenter 84.2 2.4 52.9 6.2 70.9 2.0 44.6 4.4Laptop (Wi-Fi) Fixed network 31.5 1.5 20.4 2.4 28.3 1.6 18.6 1.8Laptop (Wi-Fi) Devices 87.0 22.0 74.2 8.7 84.0 20.0 70.4 7.0Total 202.8 25.9 147.5 17.3 183.3 23.6 133.6 13.2Laptop (Mobile) Datacenter 84.2 2.4 52.9 0.7 70.9 2.0 44.6 0.9Laptop (Mobile) Fixed network 24.5 1.2 15.9 0.2 21.8 1.3 14.3 0.3Laptop (Mobile) Mobile network 223.5 66.1 200.3 2.6 43.3 12.0 37.9 0.8Laptop (Mobile) Devices 87.0 22.0 74.2 1.0 84.0 20.0 70.4 1.5Total 419.2 91.7 343.2 4.4 220.0 35.2 167.2 3.5Smartphone Datacenter 5.4 0.2 3.4 0.4 16.5 0.5 10.4 2.1Smartphone Fixed network 1.2 0.1 0.8 0.1 4.2 0.2 2.8 0.6Smartphone Mobile network 26.2 7.8 23.5 2.9 15.6 4.3 13.7 2.8Smartphone Devices 3.0 14.0 15.8 1.9 3.0 13.0 14.8 3.1Total 35.8 22.0 43.4 5.3 39.3 18.0 41.6 8.6Tablet (Wi-Fi) Datacenter 18.0 0.5 11.3 0.5 69.1 1.9 43.4 3.4Tablet (Wi-Fi) Fixed network 6.7 0.3 4.4 0.2 27.6 1.6 18.2 1.4Tablet (Wi-Fi) Devices 12.0 60.0 67.2 2.9 12.0 55.0 62.2 4.9Total 36.7 60.8 82.9 3.5 108.8 58.5 123.8 9.8Tablet (Mobile) Datacenter 18.0 0.5 11.3 0.2 69.1 1.9 43.4 1.8Tablet (Mobile) Fixed network 5.2 0.2 3.4 0.1 21.5 1.2 14.2 0.6Tablet (Mobile) Mobile network 49.1 14.5 44.0 0.8 40.0 11.0 35.0 1.4Tablet (Mobile) Devices 12.0 60.0 67.2 1.2 12.0 55.0 62.2 2.5Total 84.2 75.3 125.8 2.2 142.6 69.2 154.8 6.3TV platform Broadcast TV network 21.4 0.0 12.9 4.6 21.4 0.0 12.9 4.8TV platform Datacenter 62.4 1.8 39.2 14.0 72.3 2.0 45.4 16.9TV platform Fixed network 23.4 1.1 15.1 5.4 28.9 1.7 19.0 7.1TV platform Devices 340.0 32.0 236.0 84.5 326.0 30.0 225.6 83.8Total 447.2 34.9 303.2 108.6 448.6 33.7 302.9 112.5Total [all platforms] 172.6 179.8Table C.3: Energy and GHG emissions by platform154C.2 Model input parametersC.2.1 Device parametersOperational energy useOperational energy use is based on a Consumer Electronics Association survey based on US consumersin 2010 [5]; for tablets and smartphones which had limited coverage in that study, laboratory measure-ments are applied giving a pessimistic estimate [217]. The CEA study is an update of a previous 2007study by Roth et al [4]; trends evident across these two studies are applied in order to estimate deviceannual energy use in 2012 and 2017. All forecasts to 2017 are based only on the trends from thesestudies and discussion in [5]; details are below.Device unit electricity (kWh/year)2006 2010 2012 (est) 2017 (est) CAGR 2010-2017Desktop PC 237 220 212 193 -1.8%Laptop PC 72 63 61 56 -1.6%LCD monitor 85 97 100 109 1.7%Game console 36 135 149 190 5.0%Set top box 131 136 136 136 0TV set 249 183 173 152 -3%DVD / Blu-ray player 36 28 25 18 -6%AV receiver – 65 65 65 0Tablet – – 12 12 0Smartphone – – 3 3 0Table C.4: Device operational energy assumptions, with 2006 baseline from [4] and 2010 baselinefrom [5].Desktop PC UEC declined by 1.8%/yr from 2006 to 2010 largely due to higher availability of powermanagement features and decreases in active power draw; these trends are assumed to continue.Set-top box power grew slightly in the case of cable STBs and declined slightly for satellite STBs;average power is assumed to be stable through 2017, though this may decline if better power manage-ment features are designed.LCD monitor energy grew by 3.3%/yr from 2006 to 2010 largely due to growth in average screen155size; this growth is expected to continue slightly, but at half the rate to account for improved powerefficiency.DVD player energy decreased by 6%/yr due to improved power management features, important asthese are very low-utilization devices; this trend is assumed to continue, in part due to power manage-ment and in part due to reduced usage in favour of streaming video.Laptop energy decreased by 3.3%/yr from 2006 to 2010 due to improved power management fea-tures; as laptop energy is already quite low, it is assumed to continue declining at half this rate through2017.Game console energy grew by 40%/yr from 2006 to 2010 due to large changes in device form andfunction, as these devices found a niche as media access platforms for connected TVs; active power perconsole spiked in 2006 with the introduction of a new generation of consoles and then quickly declinedas more efficient versions were introduced. An NRDC blog post estimates the new XBox One and PS4consume 253 and 184 kWh/year, respectively [245] – significantly higher than the CEA’s average of135 kWh/yr in 2010, though these numbers are expected to decline as subsequent models become moreefficient. Overall, a modest 5%/yr gain in annual energy is assumed from 2010 to 2017.AV receivers were not modeled in the 2007 study; their energy consumption is assumed to remainstable through 2017.TV energy declined by 7.5%/yr from 2006 to 2010, but only by 2.7%/yr from 2009 to 2010. Com-peting trends are at play; growth in new TV screen size creates upward pressure on power consumption,but this is more than offset by large gains in LCD energy efficiency. LCDs and plasma TVs whichentered the market from 2006 to 2009 had much higher active power consumption than LCDs manufac-tured in 2010. Accordingly, the 2.7%/yr decline in energy consumption is assumed to increase through2017 as more efficient LCDs achieve greater market share.Smartphone and tablet energy is not covered in these two studies; energy consumption is assumedto remain stable through 2017 as these devices are already heavily power-optimized.Embodied emissionsDevice embodied GHG emissions per product GD,emb. and device lifespan LD, which together determineannual device embodied GHG emissions, are based on a comprehensive review of prior LCA literaturedeveloped by researchers at LBNL [42]. For forecasts to 2017, the US industrial average decline of1561.75% per year is applied, per Energy Star, so that embodied emissions per device decline by 9.2% from2012 to 2017.Installed baseInstalled base is used in the model to calibrate device traffic rates and to scale device impacts to obtainthe average person’s impacts due to ICT end-uses. Baseline estimates for device installed base areobtained from Consumer Electronics Association survey data for US consumers in 2010 [5]. For TVsets, set-top boxes, and game consoles, household penetration rates are assumed to be constant through2017; installed base estimates for 2012 and 2017 are generated by scaling 2010 data according to USpopulation growth. Because PC, tablet, and smartphone ownership rates are in flux, more detailedforecasts were obtained for these devices.Tablet installed base is based on a forecast from Forrester [111] which estimated 60MM tablets in-use in a consumer context in the US in 2012; extrapolating trends from 2016 to 2017 gives an estimateof 120MM tablets in 2017. This estimate is exclusive of tablets used for work purposes. Smartphone in-stalled base is based on a forecast from eMarketer [112] which estimated there were 120M smartphoneusers in the US in 2012, growing to 207MM in 2017; we assume a one-to-one mapping between smart-phone users and smartphone devices in use. Installed base forecasts for desktop and laptop PCs wereconstructed from shipment forecasts from IDC [218], assuming 20% annual retirement for desktops and25% for laptops, respectively corresponding to 5 and 4 year lifespans. IDC forecasts a 22% decline inannual desktop shipments and a 10% decline in laptop shipments from 2012 to 2017, which translates toa 12% decline in desktop installed base and an 8% decline in laptop installed base over that time span.LCD monitor installed base is assumed to follow desktop and laptop installed base given reportedratios of 0.96 monitors per desktop and 0.26 monitors per laptop [5].C.2.2 Network and datacenter intensitiesEnergy and embodied GHG intensities are obtained by dividing total energy and total embodied emis-sions by total data traffic. All traffic estimates are obtained from the Cisco VNI [102], for a US consumercontext unless otherwise specified. Cisco estimates US consumer traffic to be 135.8 EB/yr in 2012, ofwhich 134 EB/yr travels by fixed network and 1.8 EB/yr on mobile networks; in 2017, consumer trafficis 386.5 EB/yr, with 369 EB/yr fixed and 17.5 EB/yr mobile. Consumer traffic represents 86% of total157US IP traffic in 2012 and 87% of total IP traffic in 2017. Including both consumer and business, US IPtraffic represents 31% of global traffic in 2012 (23% of mobile and 31% of fixed), and 31% of globaltraffic in 2017 (18% of mobile and 32% of fixed).Fixed and mobile networks are each modeled as a combination of fixed or mobile access networkplus a part of a core network. To obtain fixed and mobile GHG intensities, core network GHG intensityand access network GHG intensity are obtained separately and then summed.Core IP network intensityCore IP network estimates are based on work from Malmodin et al, who conducted a life cycle assess-ment of shared data transport/transmission and IP edge/metro/core networks in Sweden [246]; that studyestimated 126 GWh electricity and 22 kt CO2e for Swedish core IP networks in 2010 carrying 1.5 EBof data, which yields an energy intensity of 0.08 kWh/GB and an embodied GHG intensity of 0.014kg CO2e/GB. This estimate is applied directly assuming that US core IP networks are comparable toSwedish networks. Swedish core IP network energy grew 4.6% annually from 2006 to 2010 [246];this rate of growth is assumed to continue through 2017. As global data traffic doubled from 2010 to2012 [102], core IP network energy intensity would thus decline to 0.04 kWh/GB in 2012 and 0.01kWh/GB in 2017; likewise, assuming similar trends for embodied emissions, embodied GHG intensitydeclines to 0.6 g CO2e/GB in 2012 and 0.3 g CO2e/GB in 2017.One other data source is Schien et al [247] which estimated between 0.010 and 0.023 kWh/GBfor core networks using a bottom-up model, though a later study by the same authors [41] includesMalmodin’s result of 0.08 kWh/GB as an upper bound. The CEET wireless cloud study assumed43µJ/bit [43], based on a bottom-up model from Baliga et al [248], which corresponds to about0.1kWh/GB.Mobile access network intensityMobile access network intensity is based on a 2011 forecast that estimates radio access networks glob-ally use of 72 TWh/yr energy in 2012 and 91 TWh/yr in 2017, with corresponding embodied emissionsof 22 Mt CO2e/yr in 2012 and 26 Mt CO2e/yr in 2017 [249]; these results were interpolated from thestudy’s original forecast for 2007, 2014, and 2020. The forecast is inclusive of all deployed networksglobally including 2G, 3G, and subsequent generations, and accounts for likely continuous improve-158ments in energy efficient technology. An independent study from the GSM Association estimates thatglobal mobile networks consumed about 87 TWh of electricity in 2010 (the original study reports 78TWh electricity and 43 TWh primary energy from diesel generators, of which 9 TWh is output as elec-tricity) [250]. This estimate is about 40% higher than the forecast we have applied, but includes a shareof core IP network energy on top of global radio access network energy; the rough agreement givessome confidence in their accuracy.Based on global mobile data forecasts of 10.7 EB/year in 2012 and 132 EB/year in 2017 fromthe Cisco VNI[102], mobile access networks have a global energy intensity of 6.7 kWh/GB in 2012and 0.7 kWh/GB in 2017, and a global embodied GHG intensity of 2.1 kg CO2e/GB in 2012 and 0.2kg CO2e/GB in 2017. Cisco reports that US mobile traffic is about 23% of global mobile traffic in2012 and will account for 18% in 2017. The GSM study reports that USA and Canada are collectivelyresponsible for 14.2% of global mobile network energy in 2010 [250]; if the US is assumed to accountfor 13% of global mobile network energy and embodied emissions in 2010 (subtracting out Canada’senergy), and this result is assumed to be stable to 2012, then US mobile networks produce 23% ofglobal mobile traffic using 13% of global mobile access network energy, implying that US mobileaccess networks are 40% less energy intense than the global average. Assuming that this is true forembodied emissions as well, and that this ratio holds into 2017, this yields US mobile network accessenergy intensity of 3.7 kWh/GB in 2012 and 0.4 kWh/GB in 2017, and embodied GHG intensity of1.1 kg CO2e/GB in 2012 and 0.1 kg CO2e/GB in 2017.Mobile network overall intensityWhen summing mobile access network intensity and core IP network intensity, the latter is negligible;combined mobile network intensity is equivalent to mobile access network intensity, i.e. energy intensityis 3.7 kWh/GB in 2012 and 0.4 kWh/GB in 2017, while embodied GHG intensity is 1.1 kg CO2e/GBin 2012 and 0.1 kg CO2e/GB in 2017.By comparison, Malmodin estimates 3G mobile broadband networks have an energy intensity of 2.9kWh/GB, while 2G and PSTN have energy intensities of 37 kWh/GB and 18 kWh/GB, respectively,considering a Swedish context in 2010 [251]. The majority of traffic in such a system would be carriedby the more modern 3G network, but energy consumed by older networks will push up the averageenergy intensity. The estimates we have applied, derived from [249], appear to be consistent with159Malmodin’s.The CEET wireless cloud study [43] applied an estimate that 4G LTE wireless technology usesbetween 328 uJ/bit and 615 uJ/bit in 2010 [252], improving 26% per year [219]; this would correspondto 0.40 to 0.75 kWh/GB in 2012 and 0.088 to 0.17 kWh/GB in 2017. Our estimate for overall mobilenetwork energy intensity for 2017 is significantly higher than the forecasted energy intensity of 4G LTEalone, but the study our forecast is based upon also includes the energy of older 3G networks whichwould carry some of the traffic.Schien et al [41] applied a distribution of estimates for mobile networks ranging from 0.030 kWh/GB(min), 0.12 kWh/GB (mode), and 0.73 kWh/GB (max), modeling 3G networks in the UK in 2012, all ofwhich are several times lower than our estimates, though the first two are based on a bottom-up model.The latter figure, 0.73 kWh/GB, was obtained by dividing total base station power by total traffic. Theyused a Vodafone estimate of 2.1 kW per base station, and having obtained an energy intensity of 328J/Mb (0.73 kWh/GB), thus assumed average base station traffic of 6.5 Mb/s. Our forecast is built upona study [249] that assumes a global average of 1.3 kW per base station and an average traffic rate of1.5 Mb/s per base station in 2014, yielding an energy intensity of 870 J/Mb (1.9 kWh/GB); when thatstudy’s results are interpolated back to 2012 the intensity grows larger. The total global mobile trafficestimate used in [249], 45 EB/yr in 2014, is slightly larger than Cisco’s estimate of 31 EB/yr in 2014 wewhich have applied. The discrepancy between our forecast and the result used by Schien et al appearsto be related to the assumption of average mobile data traffic; average traffic per base station in [249]may be lower due to the inclusion of older technologies like 2G, while Schien et al focused on 3G only.Fixed access network intensityFixed access network energy is based on a study from Lanzisera et al which estimates that customeraccess and residential customer premise equipment collectively consumed 10.5 TWh/yr in the US in2012, having grown 11% per year since 2007 [208]. The study was based on an estimate of US installedstock alongside laboratory power measurements; its authors estimate total results to be within about20% of the nominal estimate, based on a basic sensitivity analysis. In the absence of any availableforecast, we have assumed the same trends continue, such that fixed access networks will consume 18TWh/yr in the US in 2017. The majority of fixed access network energy is due to customer premiseequipment, i.e. modems and routers; according to Consumer Electronics Association data, broadband160network devices account for 6.3 TWh/yr in US residences in 2010 [5], which is roughly comparable tothough slightly lower than the results from Lanzisera [208]. With US fixed consumer traffic being 134EB/yr in 2012 and 369 EB/yr, fixed access network energy intensity is therefore estimated to be 0.08kWh/GB in 2012, and 0.05 kWh/GB in 2017.This estimate is much lower than Malmodin’s assumption of 0.3 kWh/GB for modems and routers,plus 0.08 kWh/Gb for access lines [251] in 2010. To some degree this can be explained by reducedenergy intensity caused by a doubling of data traffic from 2010 to 2012, as Malmodin’s study assumeda data year of 2010. In addition, their estimate for the the energy consumption of network devices –118 kWh/yr per household, accounting for 1.5 devices per household – differs from CEA data for USconsumers in 2010 which estimated 1.2 devices per household, with unit energy consumption around50 kWh/yr per device [5]. This implies about half of the energy consumption as Malmodin’s study, butwith double the traffic, accounting for the roughly 4X difference in energy intensity estimates.There are no direct studies of embodied emissions of fixed access equipment in the US, but a study ofglobal networks estimated that embodied GHGs for customer premise equipment, the main componentin fixed access networks, was estimated to be 2.3 Mt CO2e/yr in 2007 for equipment using 35 TWh/yrof electricity [25], which yields a ratio of 65 kt CO2e/TWh. This is again much smaller than emissionsdue to use electricity, which would be roughly 600 kt CO2e/TWh assuming a US average grid. Thisratio of embodied emissions to use energy was applied to the previous estimates, producing first-orderestimates of 0.7 Mt CO2e/yr in 2012 and 1.2 Mt CO2e/yr in 2017.Fixed network overall intensityBy summing fixed access network and core network intensity, overall fixed network intensity is esti-mated to be 0.12 kWh/GB in 2012 and 0.06 kWh/GB in 2017, while embodied GHG intensity is 0.006kg CO2e/GB in 2012 and 0.003 kg CO2e/GB in 2017.This may be compared to a pessimistic bottom-up study of a fixed link videoconference which esti-mates an upper bound of 0.2 kWh/Gb for average data transfers in 2009 [221]; with improved efficiencysince then, this would likely be below our estimate for 2012; however, that study applied a bottom-upmethodology which does tend to produce lower estimates than top-down approaches. Other estimates offixed network intensity reviewed in [72] are from older data years. For comparison to additional studies,see the discussion on core network intensity and fixed access network intensity above.161Datacenter intensityDatacenter energy is taken from a recent industry survey by DatacenterDynamics, which estimated a to-tal of 87 TWh in the US in 2012, inclusive of all cooling and infrastructure energy [253]. This estimateis comparable to Koomey’s estimate of 76 TWh in 2010 [23] and was generated independently. Data-center Dynamics forecasts North American datacenter energy will reach 116 TWh/yr in 2016, growing6% annually; from this we estimate US datacenter energy to be 116 TWh/yr in 2017. The forecastannual growth rate of 6% matches the growth rate observed by Koomey from 2005 to 2010, a period oftime which included an economic downturn.Total US IP traffic in 2012 is 158 EB/year according to Cisco [102], which implies an aggregateUS datacenter energy intensity of 0.55 kWh/GB of IP traffic. However, while about 86% of IP trafficis generated due to consumer end-uses according to Cisco, a significant portion of servers are servicingbusiness workloads. A study of US business end-uses estimated that about 80% of server closet, serverroom, and localized datacenter energy is consumed serving business applications [42]; those datacenterswere collectively responsible for about 60% of total datacenter energy in 2008. For the remaining mid-tier and enterprise-class datacenters, we assume 80% of their energy consumed is due to consumerend-uses, on the basis that 80% of inbound traffic to these data centers is due to consumer end-usesaccording to Cisco [108]. Therefore, we assume about 50% of datacenter energy is consumed in servingconsumer workloads, i.e. 43 TWh/yr in 2012, and 58 TWh/yr in 2017. This corresponds to datacenterenergy intensity of 0.32 kWh/GB in 2012 and 0.15 kWh/GB in 2017, or 328 J/Mb in 2012 and 866J/Mb in 2017.Unfortunately, this approach has a major discrepancy when compared with other studies; Schienet al report bottom-up energy intensity of 0.89 J/Mb for Akamai CDN servers [41]; Chandaria et alassumed 0.4 J/Mb for CDN servers streaming video content from BBC [45]. The CEET wireless cloudstudy [43] estimated 20 J/Mb for servers, based on data from Facebook and Google. On the otherhand, a top-down model from Malmodin estimated 1 kWh/GB for data centers in 2010 using a similarmethodology to our study [251].Google estimates that an active user performs 25 searches and watches 60 minutes of Youtubecontent per day and calculates a corresponding daily footprint of 8g CO2e/day [224]. Google reportstheir carbon intensity to be 357 g CO2e/kWh – better than the US grid average – implying that this162activity required 0.022 kWh of energy. Assuming this activity corresponds to roughly 0.3 GB of datatraffic, this corresponds to an energy intensity of 0.075 kWh/GB, or 33 J/Mb – larger than the bottom-upestimates above. Google also indicates that their datacenters are much more efficient than the industryaverage. Applying Google’s energy to total US consumer traffic levels of 136 EB/year in 2012 wouldyield a total energy consumption of 10.2 TWh – less than 25% of total estimated datacenter energyconsumption due to consumer workloads.If Google’s data centers are indeed among the most efficient, the industry-wide average datacenterintensity should be above 0.075 kWh/GB. However, our 2012 estimate of 0.32 kWh/Gb is four timeslarger and could conceivably be too large, especially since our estimation of the total amount of dat-acenter energy dedicated to consumer workloads – 50% of the total – is based on rough assumptions.Nevertheless, if datacenter energy intensity was lowered to be closer to Google’s reported intensity,this would leave a large portion of total datacenter energy unallocated. Better data is needed in orderto identify the portion of datacenter energy dedicated to consumer workloads; ideally a disaggregatedmodel could be developed to account for different levels of energy intensity which correspond to dif-ferent network services. In the mean time, our study proceeds under the above original assumptionswith datacenter energy intensity of 0.32 kWh/GB in 2012 and 0.15 kWh/GB in 2017, with a sensitivityanalysis to test the effects of applying Google’s intensity as the industry average.Embodied GHG emissions are estimated based on a 2008 inventory of the total number and typeof servers in US datacenters [24] and the approximate embodied emissions of each [42]; this yieldeda ratio of about 30 kT CO2e/TWh of use-phase electricity in 2008, which we assume scales linearlyas use electricity grows. This estimate is first-order, but since use emissions would generate about600 kT CO2e/TWh, embodied GHG emissions likely contributes less than 10% to overall datacenteremissions and do not warrant deeper investigation. Datacenter embodied GHG intensities are estimatedto be 0.009kg CO2e/GB in 2012 and 0.004kg CO2e/GB in 2017.C.2.3 End-use parametersData traffic per end useTypical device traffic per month estimates are derived from the Cisco VNI and related reports[102, 103,254]. The study provides estimates of monthly traffic per device among mobile-connected devices;163share of traffic offloaded to fixed networks from mobile devices; installed base of mobile devices; andoverall network traffic due to TVs, PCs, tablets, and smartphones. Using our own estimates of PC andTV installed base, we derive average monthly traffic estimates for each of the four devices, the resultsof which are in the main text. In order to confirm that the Cisco model has been interpreted correctlyand that our installed base estimates are consistent with the internal assumptions in the Cisco model,overall traffic due to each type of device was calculated and is shown in Table C.5; at the bottom, totalUS consumer traffic due to all devices in our model is compared to total US consumer traffic as reportedby Cisco. The numbers agree within 2% for 2012, suggesting that our assumptions are reasonable; aslight gap in traffic in 2017 is likely due to additional devices not modeled here but included in the Ciscomodel, such as smart glasses and machine-to-machine modules.Installed base (millions) Fixed traffic (EB/yr) Mobile traffic (EB/yr)2012 2017 2012 2017 2012 2017TVs 358 371 70 178PCs, mobile-connected 13 21 3 8 0.8 2.3PCs, fixed/wifi only 219 189 58 89Tablets, mobile-connected 17 41 1 15 0.2 4.1Tablets, fixed/wifi only 43 79 2 36Smartphones 121 207 1 15 0.8 8.2All devices total (EB/year, this study) 135 341 1.8 14.6US consumer total (EB/year, via Cisco) 134 369 1.8 17.5Table C.5: Total traffic from all devices, comparing this study assumptions with original Ciscomodel.The Cisco VNI stratifies network traffic by major end-use. Considering internet traffic which isexclusive of IP video-on-demand, fixed network traffic is 68% online video in 2012 and 75% in 2017,while mobile network traffic is 50% online video in 2012 and 62% in 2017. Considering overall IP trafficwhich includes internet and IP video-on-demand, IP video-on-demand represents 45% of IP traffic in2012 and 42% of IP traffic in 2017, all of which travels on fixed networks.The distribution of network traffic by end-use is not known at a device level, e.g. it is unknownwhat percentage of a desktop PC’s fixed internet traffic is attributable to video, with the exception ofIP video-on-demand to televisions which is explicitly reported; of the average monthly TV traffic per164device, 14.5 GB/month in 2012 and 33.4 GB/month in 2017 is due to IP video on demand. For all otherdevices and networks, the share of online video traffic is assumed to match the percentages listed above,i.e. fixed traffic is 68% video in 2012 and 75% online video for all devices; mobile traffic is 50% onlinevideo in 2012 and 62% online video in 2017 for all devices that use mobile networks.Derivation of time spent with online videoTime spent on each end-use on each device is estimated using the above traffic figures alongside variousmarket research reports which track consumer behavior. This category is difficult to accurately estimatebecause of large amounts of variation across different market research reports, likely due to differencesin scope and boundary or modeling assumptions, along with poor coverage in academic literature. Theapproach taken here is to estimate average online video data rates, in GB per hour, so that time spentwith online video on each device can be calculated directly from the previously derived traffic model;data rate estimates are tuned slightly so that calculated time spent with online video is consistent withavailable market research data, where possible. Data rate assumptions and resulting estimated timespent per person in the US are shown in Table C.6. The data rates may be compared with published datarates of Netflix online streaming: 0.3 GB/hr at low quality, 0.7GB/hr at medium quality, 1.0 GB/hr athigh quality at standard definition, and 2.8 GB/hr at high quality at high definition [255].Video traffic(GB/monthper device)Devicepenetration (%of population)Video trafficrate (GB/hr)Online videotime spent(hrs/mon-th/person)2012 2017 2012 2017 2012 2017 2012 2017PC 14.8 29.6 0.74 0.67 1 1.5 11.0 13.2Tablet 3.1 29.1 0.19 0.37 0.3 0.8 2.0 13.4Smartphone 0.7 7.3 0.39 0.64 0.3 0.8 0.9 5.8TV (internet) 1.8 6.4 1.15 1.15 0.8 1.2 2.6 6.2TV (IPVoD) 14.5 33.4 1.15 1.15 0.8 1.2 20.8 31.9Total (internet and IPVoD) 37.3 70.4Total (internet video only) 16.5 38.6Table C.6: Derivation of time spent with online video.The model’s estimate of 16.5 hours of online video per month per person in 2012 (excluding IP165video on demand) is consistent with eMarketer’s estimate of 16.1 hrs/month per person on PC, tablet,and smartphone only [209] (of which about 2 hrs/month per person occurs on tablets and 2 hrs/monthper person on smartphones), and similar to an Alcatel-Lucent study which relied a model similar toCisco’s and estimates 16.2 hrs/month per person [100]. Other estimates for 2012 are 13.6 hrs/monthfrom comScore [101] and only 4.2 hrs/month per person from Nielsen [98]. It is difficult to reconcilethe low Nielsen estimate with the levels of online video traffic estimated by Cisco, as that would implyunrealistically high average data rates; this interesting discrepancy is unresolved.Alcatel-Lucent predicts that online video will grow to 102 hrs/month by 2017 accompanied by asevere 80% decline in traditional TV consumption [100], but we are hesitant to adapt such a drasticforecast here, especially given Nielsen’s recent reports which suggest that time spent watching tradi-tional TV has been relatively stable into 2013 [98]; likewise this forecast is incompatible with Cisco’sdata forecast without an unrealistically large increase in video data rates. Forecasts in the Alcatel-Lucentstudy show data rates for online video services growing roughly 50% from 2012 to 2020. We have as-sumed comparable levels of growth in data rates, which, given the data forecast from Cisco, leads to adoubling in total time spent with online video per person from 2012 to 2017, with most of the growthoccurring on tablets and TV sets. Traditional TV is estimated to be 138 hrs/month per person in 2012based on Nielsen data, and is assumed to decline by 10% to 124 hrs/month in 2017.Time spent with other end-usesEach device is actively used for a certain number of hours per month, with that time split amongst onlinevideo, online non-video, offline, and traditional television. Some data sources track the amount of timespent online, either on specific devices or per person overall; others track the amount of time spent onspecific devices. Due to differences in methodology across studies it is difficult to compare results; ourmodel is developed with the goal of being consistent with the majority of data sources, as well as withthe online video time spent estimates previously derived. Several studies track how time spent online isdistributed across devices, which makes a good benchmark for validation; these are listed in Table C.7below, along with our assumptions for 2012 and 2017. Each of the three studies compared against agreesthat the majority of time spent online happens on PCs; the distribution across tablet and mobile devicesvaries according to assumptions of tablet penetration, which grew significantly from 2012 to 2013.The estimates were based on a few key assumptions, as follows. Tablets devices are each used for166eMarketer(2013)[209]GfK (2012)[256]comScore(2013) [101]This studyassumptions2012 2017PC 49% 73% 52% 67% 42%Tablet 39% 6% 13% 10% 19%Mobile 12% 17% 35% 21% 37%TV 4% 2% 2%Total hours/month 71 127Table C.7: Modal share of time spent online per person.about 60 hours/month, per eMarketer [209]. The average US consumer in 2012 spent 20 hours/month onsmartphone non-voice activities according to eMarketer; voice activities account for 20% of smartphonetime according to GSMA Intelligence [257]; and about 40% of US consumers owned a smartphonein 2012 according to our estimates; therefore the average smartphone device was used for about 65hours/month in 2012. The average US person spent 75 hours/month online on PCs in 2012 accordingto eMarketer[209]; 64% of PC online time occurs in the home, according to Temkin Group [258]; andthere are 0.74 PCs in use per person in the US; therefore the average PC is used for 64 hours/month foronline uses only. Among French users in 2008, offline PC time comprised 38% of the total [259]; thishas likely declined; we assume 30% of PC time is offline in 2012, so that the average PC is used fora total of 92 hours/month. Tablets and smartphones are each assumed to be used for online end-uses60% of the time in 2012, with the remaining time spent on gaming, phone calls and messaging, andother uses. For all devices, total hours spent online is assumed to grow by 20% from 2012 to 2017; timespent offline is assumed to decline to be 10% of total device time. The resulting total amount of timespent online, 71 hours/month, is in the midrange of other estimates: Ofcom estimates UK adults spent35 hours/month online in 2012 [260]; eMartketer estimates typical online usage of 75 hrs/month on PCs(apparently including time spent at work); and Temkin Group estimates 116 hrs/month per person onPCs at home [258].Time spent online (excluding video) and offline represents a small percentage of total time spent onTVs. Nielsen reports 2.9 hours/month per person spent using DVD/Blu Ray devices, and 2.2 hours/-month per person using game consoles [98]; we assuming half of game console time is spent online,with the remainder offline.167C.3 Uncertainty modelThere are many forms of uncertainty which affect the model. All model parameters are estimates ofempirical quantities, usually estimating a US consumer average, and have varying degrees of precision.Very few of the model parameters used in our study include any treatment of uncertainty or expression ofconfidence in their source studies. Nevertheless, a semi-quantitative treatment uncertainty is undertakenby assigning 95% confidence intervals for each parameter according to our best judgment, supported byevidence where available.An additional large source of uncertainty is structural; any number of the modeling assumptionscould be challenged, and a change in parameters could greatly affect the outputs. For example, if ourassumption for average datacenter energy intensity was reduced to be closer to that reported by Google,then datacenter energy tabulated for each end-use would drop accordingly – perhaps being cut in half.There is little utility in assigning confidence intervals that span the entire range of possible modelingassumptions. Rather, the impact of changes in modeling assumptions would be more appropriately as-sessed through scenario analysis, as was performed in the text for the case of datacenter energy intensity.Each modeling assumption is reasonable according to our best judgment and is thoroughly justified, andthe model has been specified in such a way as to make it easily adaptable should other researcherswish to explore the impacts of changes in modeling assumptions. Therefore, the uncertainty analy-sis assumes that the underlying modeling assumptions are fixed, and attempts to calculate reasonableconfidence bounds on the outputs under these assumptions.Output model uncertainty is estimated through Monte Carlo analysis, by assuming normally dis-tributed inputs, such that the assigned confidence intervals represent a range of ±1.96 standard devia-tions from the mean.C.3.1 Parameter uncertainty estimationDevice energy and emissionsUncertainty bounds for device energy consumption are based on field measurements from a Minnesotaplug load study [228]. Full field data is not available, but the study reports sub-categories of devicesbased on their physical characteristics or usage profiles from which we can roughly estimate populationcharacteristics, assuming Minnesota residents are representative of US consumers. Of desktop PCs, 42168were sampled with a mean energy consumption of 282 kWh/yr, and standard deviation of 202 kWh/yr.This yield a standard error of 32 kWh/yr, implying that the larger population mean lies within ±22% ofthe sample mean with 95% confidence. Likewise, the same study measured a sample of 110 TV sets.These TV sets had a mean energy consumption of 166 kWh/yr per device, with standard deviation of176 kWh/yr, and a standard error of 17.9 kWh/yr implying that the population mean lies within ±19%of the sample mean with 95% confidence. These two products are probably the most variable amongconsumer electronics devices due to their wide range in physical forms and in usage patterns; thusbounds of ±20% could be conservatively applied to all other devices. Energy estimates for 2017 aresomewhat speculative, and are thus assigned wider bounds of ±30%.No such sample exists for embodied emissions, but a study of uncertainty in process-sum LCAof ICT equipment, using a server as a case study, estimated bounds of about ±15% for embodiedemissions [133]. There exist structural uncertainties in the estimation of embodied emissions, especiallyrelating to modeling assumptions; for example, process-sum LCA is vulnerable to cut-off error and mayunder-report emissions by 50% or more, when compared to hybrid or economic input-output LCA [2].Such structural uncertainties are beyond the scope of the uncertainty assessment of this study. Thisstudy assumes bounds of ±20% for device embodied emissions in 2012 and ±30% in 2017.Device installed base is assumed to have bounds of ±5% in 2012 and ±10% in 2017 in most cases,as it is informed by high-quality data sources. The exception is smartphones and tablets, both of whichare undergoing rapid changes in penetration due to their relative novelty. For these two devices, installedbase is assumed to have bounds of ±10% in 2012 and ±20% in 2017.Network and datacenter intensitiesNetwork and datacenter intensities each rely on a ratio of energy or embodied emissions divided by totaltraffic. The traffic data is from Cisco and is probably accurate to within ±10% in 2012, given historicalaccuracy reported by the authors [102]; 2017 traffic is more uncertain. Datacenter and mobile networkoverall energy were both corroborated by two independent studies and are assumed to be relatively ac-curate. Fixed network overall energy is dominated by customer premises equipment which is obtainedfrom high-quality consumer electronics data [5]. Overall, it seems reasonable to assume that intensitiesare accurate within ±20% in 2012 and ±40% in 2017, assuming the modeling assumptions are appro-priate. Embodied GHG intensities have poorer quality data sources and are thus assigned wider bounds169of ±30% in 2012 and ±50%.Network and datacenter intensities are the most vulnerable to structural uncertainties in the modelingassumptions; especially datacenter uncertainty due to disagreement among other published studies. Forthis reason, a scenario analysis of datacenter intensity was performed in the text.End-use traffic and timeTraffic per end-use is obtained directly from the Cisco model, which reports traffic at a disaggregatedlevel; these are assumed to be accurate within ±10% in 2012 and ±20% in 2017, based on Cisco’sreported 10% historical accuracy as discussed above. Again, as these parameters (e.g. typical monthlyvideo traffic per tablet) represent estimates of the population mean, these confidence intervals describebounds on the mean, and do not capture variability within the population.User time spent for each device was obtained from a synthesis of market research reports; estimateswere tuned to be as consistent as possible with each data source. Additional calibration for time spentwith online video was performed using the traffic model. Accordingly, these estimates are believed tobe fairly accurate; confidence intervals are assumed to be ±15% in 2012 and ±30% in 2017.170


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items