EPRI：2024為智能提供動力人工智能與數據中心能耗分析白皮書（英文版）（35頁）.pdf

《EPRI：2024為智能提供動力人工智能與數據中心能耗分析白皮書（英文版）（35頁）.pdf》由會員分享，可在線閱讀，更多相關《EPRI：2024為智能提供動力人工智能與數據中心能耗分析白皮書（英文版）（35頁）.pdf（35頁珍藏版）》請在三個皮匠報告上搜索。

1、2024 White PaperPowering IntelligenceAnalyzing Artificial Intelligence and Data Center Energy Consumption143222532|Powering Intelligence May 2024EXECUTIVE SUMMARYKey Messages In the United States,powering data centers,providing clean energy for manufacturing,supporting industrial onshoring,and elect

2、rifying transporta-tion are driving renewed electric load growth.Clusters of new,large point loads are testing the ability of electric companies to keep pace.Data centers are one of the fastest growing industries worldwide.Between 2017 and 2021,electricity used by Meta,Amazon,Microsoft,and Googlethe

3、 main providers of commercially available cloud computing and digital servicesmore than doubled.A fundamental uncertainty in projecting data center load growth comes from the broad emergence of artificial intelligence(AI)technologies in business and daily lifepunctuated by the explosion into public

4、consciousness of gen-erative AI models,such as OpenAIs ChatGPT,released in November 2022.While AI applications are estimated to use only 10%20%of data center electricity today,that percentage is growing rapidly.AI models are typically much more energy-intensive than the data retrieval,streaming,and

5、communications applications that drove data center growth over the past two decades.At 2.9 watt-hours per ChatGPT request,AI que-ries are estimated to require 10 x the electricity of traditional Google queries,which use about 0.3 watt-hours each;and emerging,computation-intensive capabilities such a

6、s image,audio,and video generation have no precedent.To provide an early assessment of potential data center load growth at the national level,EPRI has developed low,moderate,high,and higher growth scenarios for data center loads from 2023 to 2030.Data centers grow to consume 4.6%to 9.1%of U.S.elect

7、ricity generation annually by 2030 versus an estimated 4%today.While the national-level growth estimates are significant,it is even more striking to consider the geographic concentration of the industry and the local challenges this growth can create.Today,fifteen states account for 80%of the nation

8、al data center load,with data centers estimated to comprise a quarter of Virginias electric load in 2023.Concentration of demand is also evident globally,with data centers projected to make up almost one-third of Irelands total electricity demand by 2026.With the shift to cloud computing and AI,new

9、data centers are growing in size.It is not unusual to see new centers being built with capacities from 100 to 1000 megawattsroughly equivalent to the load from 80,000 to 800,000 homes.Connection lead times of one to two years,demands for highly reliable power,and requests for power from new,non-emit

10、ting generation sources can create local and regional electric supply challenges.EPRI highlights three essential strategies to support rapid data center expan-sion:1.Data center efficiency improvements and increased flexibility.2.Close coordination between data center developers and electric com-pan

11、ies regarding data center power needs,timing,and flexibility,as well as electric supplies and delivery constraints.3.Better modeling tools to plan the 510+year grid investments needed to anticipate and accommodate data center growth without negatively impacting other customers and to identify strate

12、gies for maintaining grid reliability with these large,novel demands.143222533|Powering Intelligence May 2024TABLE OF CONTENTSEXECUTIVE SUMMARY.2Key Messages.2Potential Impacts of Artificial Intelligence on Data Center Load Growth.4EPRI U.S.Data Center Load Projections.4Data Center Power Demands Are

13、 Concentrated in a Few Regions.5A Roadmap to Support Rapid Data Center Expansion.6Introduction.7Research Questions.7Data Centers in the United States.7Data Centers Primary Electricity-Consuming Hardware and Equipment.9AI and Data Center Power Consumption Insights.10Immense Volumes of Data are Being

14、Processed Daily.10History of Energy Efficiency in the Data Center Industry.11Uneven Geographic Distribution Creates Imbalance in Data Center Load.12AI Implications for Power Consumption.14Chat GPT and Other Large Language Models(LLMs).15Forecasting Data Center Load Growth to 2030.17Four Scenarios Ba

15、sed on Historical Data,Expert Insights,and Current Trends.17Energy Efficiency,Load Management and Clean Electricity Supply.18Energy-Efficient Training Algorithms.18Energy-Efficient Hardware.19Energy-Efficient Cooling Technologies.19Scalable Clean Energy Use.20Monitoring and Analytics.20Reducing Data

16、 Centers Environmental Footprint.21Actions to Support Rapid Data Center Expansion.21Improve Data Center Operational Efficiency and Flexibility.22Increase Collaboration through a Shared Energy Economy Model for Sustainable Data Centers.22Better Anticipate Future Point Load Growth through Improved For

17、ecasting and Modeling.23Appendix A:State-Specific Scenarios.24Projected Data Center Load Scenarios for Top 15 States.24Regional Differences in Data Center Capacities by Metropolitan Area.27Projections of Potential Power Consumption for 44 States.28Appendix B:Insights Into the Energy Use of AI Models

18、.29References.31143222534|Powering Intelligence May 2024Potential Impacts of Artificial Intelligence on Data Center Load GrowthData center operation is one of the fastest growing indus-tries worldwide.The International Energy Agency recently projected that global data center electricity demand will

19、more than double by 2026.In the United States,the national outlook could resemble the global outlook,but is highly uncertain.One key uncertainty that could change the trajectory of data center load growth is the use of generative AI models.Both public and corporate imaginations were triggered by the

20、 release of OpenAIs ChatGPT on November 30,2022.Ev-idence about how widely these tools will be used and how much they will change computational needs is just starting to emerge.These early applications were estimated to re-quire about ten times the electricityfrom 0.3 watt-hours for a traditional Go

21、ogle search to 2.9 watt-hours for a Chat-GPT queryto respond to user queries.Creation of original music,photos,and videos based upon user prompts and other emerging AI applications could require much more power.With 5.3 billion global internet users,widespread adoption of these tools could potential

22、ly lead to a step change in power requirements.On the other hand,his-tory has shown that demand for increased processing has largely been offset by data center efficiency gains.EPRI U.S.Data Center Load ProjectionsDrawing on public information about existing data centers,public estimates of industry

23、 growth,and recent electricity demand forecasts by industry experts,EPRI prepared four scenarios of potential electricity consumption in U.S.data centers during the period from 2023 to 2030(Figure ES-1).The blue line in the figure,running from 2000 to 2020,traces historical data center electricity c

24、onsumption esti-mates.From 2000 to 2010,data center load grew as centers expanded across the country to support the emerging internet.From 2010 to 2017,despite continued growth in computing demands and data storage this load growth flat-tened due to efficiency gains and the replacement of small,rela

25、tively inefficient corporate data centers with large,cloud computing facilities.In recent years,load growth has likely accelerated,driven by emerging AI applications and COVID-era increases in demand for services like streaming and video conferencing.The light blue area highlights un-certainty in a

26、range of data center electricity consumption estimates for 2021 to 2023.Colored bands show the four projections,which combine estimates of increased data processing needs with assumptions about efficiency gains.The widths of these bands carry forward the uncertainty about the 2023 starting load leve

27、l:Low growth3.7%annual load growth based on a Statista projection of data center financial growth is-sued prior to the release of ChatGPT.Figure ES-1.Projections of potential electricity consumption by U.S.data centers:20232030.%of 2030 electricity consumption projections assume that all other(non-d

28、ata center)load increases at 1%annually.6005004003002001000Electricity Consumption(TWh/y)2000200520102015202020252030MaxMinAverage historical data4.6%5.0%6.8%9.1%3.7%5%10%15%Low growthModerate growthHigh growthHigher growthSCENARIOANNUALGROWTH RATE%OF 2030 ELECTRICITYCONSUMPTION143222535|Powering In

29、telligence May 2024 Moderate growth5%annual load growth based on an expert assessment commissioned by EPRI.High growth10%annual load growth consistent with both a McKinsey estimate and another expert assess-ment commissioned by EPRI in summer 2023.Higher growth15%annual growth based upon a commissio

30、ned expert assessment consistent with rapid expansion of AI applications and limited efficiency gains.The estimates of data centers share of total U.S.electricity consumption in 20309.1%,6.8%,5.0%,and 4.6%as-sume that all other loads increase at 1%per year.Data centers accounted for about 4%of the t

31、otal load in 2023(average estimate).Data Center Power Demands Are Concentrated in a Few RegionsFifteen states accounted for an estimated 80%of the national data center load in 2023.Ranked from highest to lowest,they are Virginia,Texas,California,Illinois,Oregon,Arizona,Iowa,Georgia,Washington,Pennsy

32、lvania,New York,New Jersey,Nebraska,North Dakota,and Nevada.Concentration of demand is also evident globally,with the International Energy Agency recently projecting that data centers in Ireland could account for almost one-third of Irelands total electricity demand by 2026.The map in Figure ES-2 sh

33、ows the effect in 2030 of applying the annual U.S.data center growth rates(averaged across the four scenarios)to project state-level loads against a backdrop of 1%annual growth in other loads.With evenly spread growth,the data center share of load in Virginia in-creases to almost 50%in the higher gr

34、owth scenario and to 36%when averaged across the four scenarios.The shares in other states vary widely with five other states projected to approach 20%or more of electricity demand under these simplified assumptions.In reality,load growth is unlikely to be spread evenly.Data centers favor sites wher

35、e inter-net connections are strong;where electricity prices,land costs,and disruptive events are low;where skilled labor is available;near population centers and users;and where the centers can develop backup power to ensure power supply(usually natural gas or diesel generators).The additional Figur

36、e ES-2.2030 projected data center share of electricity consumption(assumes average of the four growth scenarios and that non-data center loads grow at 1%annually)4,8,905%510%1015%1520%20+%2030 Data Center%of State Electricity Consumption143222536|Powering Intelligence May 2024requirement of some dev

37、elopers for new,clean electricity generation sources adds to the challenge of developing and delivering this new generation.A Roadmap to Support Rapid Data Center ExpansionThe most serious challenges to data center expansion are local and regional and result from the scale of the centers themselves

38、and mismatches in infrastructure timing.A typical new data center of 100 to 1000 megawatts repre-sents a load equal to that of a new neighborhood of 80,000 to 800,000 average homes.While neighborhoods require many years to plan and build,data centers can be devel-oped and connected to the internet i

39、n one to two years.New transmission,in contrast,takes four or more years to plan,permit,and construct.And developing and connecting new generation can also take years.EPRI highlights three essential strategies to support rapid data center expansion.These strategies emphasize in-creased collaboration

40、 between data center developers and electric companies.1.Improve data center operational efficiency and flex-ibility.Although gains in data center operational ef-ficiency have plateaued in recent years,there are clear opportunities for further improvement,including more efficient IT hardware;lower e

41、lectricity use for cooling,lighting,and security;and more efficient AI develop-ment and deployment strategies.Efforts to increase both temporal and spatial(i.e.,spreading compute geographically)flexibility are critical to helping accom-modate these new loads.2.Increase collaboration between data cen

42、ter devel-opers and electric companies.Developing a deeper understanding of data center power needs,timing,and potential flexibilitieswhile assessing how they match available electric supplies and delivery constraintscan create workable solutions for all.Enabled by technol-ogy and supporting policie

43、s,data center backup gen-erators,powered by clean fuels,could support a more reliable grid while reducing the cost of data center op-eration.Shifting the data center-grid relationship from the current“passive load”model to a collaborative“shared energy economy”with grid resources power-ing data cent

44、ers and data center backup resources contributing to grid reliability and flexibilitycould not only help electric companies contend with the explo-sive growth of AI but also contribute to affordability and reliability for all electricity users.3.Improve point load forecasting to better anticipate fu

45、ture point load growth and modeling of transient system behavior to maintain reliability.Forecasts need to make better projections describing new point load locations,magnitudes,and timing alongside better techniques for making decisionsto build or not build long lead-time infrastructurewhile facing

46、 the eco-nomic,regulatory,and political uncertainty associated with siting these large point loads.Also,real-time mod-eling of data center operational characteristics in an increasingly inverter-based grid is needed to maintain reliability.143222537|Powering Intelligence May 2024INTRODUCTIONResearch

47、 QuestionsAs the number and size of data centers expand to support continued growth in data processing,internet traffic,and rapid expansion in artificial intelligence(AI)applications,some critical questions emerge:How rapidly can we expect data centers to expand,and how does the rapid growth in AI c

48、hange their power demands?What is the impact of these developments on electric load and resource adequacy?What implications do these trends have for future elec-tricity infrastructure planning?How can the data center and electric utility industries work together to support rapid data center expansio

49、n?Data Centers in the United StatesAs of March 2024,there were approximately 10,655 data centers globally;half of them,5,381,were in the United States.Just over three years ago,in January 2021,there were approximately 8,000 data centers,with about one-third of them in the United States 1.The constru

50、ction of new data centers is accelerating at a rapid pace,largely driven by demand for AI-powered tasks such as speech recognition,tailored diagnostics,logistics,internet of things(IoT),and generative AI.The expansion of interest in generative AI is particularly notable due to the overnight populari

51、ty of ChatGPT,released on November 30,2022,marking the public-facing start of a technology race.Data centers vary significantly in design and purpose and are generally grouped into two categories,small or large scale.Small-scale data centers,representing about 10%of U.S.data center load 2,typically

52、cater to localized opera-tions and service small businesses,government facilities,or specific departmental needs within larger corporations.They include server rooms/closets embedded in buildings and“edge data centers,”which are strategically located on the outer edges of networks to bring computing

53、 capabilities closer to users who are geographically distant from large cloud data centers 3.Though the electricity demands of each installation are relatively modest500 kilowatts(kW)to 2 megawatts(MW)they account for roughly half of all servers 3.Market research analysts have projected the global e

54、dge data center market to grow at a compound an-nual growth rate(CAGR)of 22.1%to 2030 4,highlighting the rising importance of small-scale and edge data centers in digital infrastructure.Large-scale commercial data centers are designed to serve extensive operations and often serve multiple businesses

55、 or even entire industries.These data centers seek proximity to customers and a skilled workforce and can benefit from lower land costs,property taxes,labor rates,energy prices,and risk of severe weather or seismic activity 5.Figures 13 show maps of various large-scale facility types,which include:E

56、nterprise data centers,which are owned and operated by single companies for their exclusive computing and networking use.These account for about 2030%of total load 2,6.Co-location centers,where several businesses may rent space to house their servers and other hardware with shared energy and cooling

57、 infrastructure.Hyperscale data centers,which are capable of rapidly scaling up their operations to meet the vast computing needs of cloud giants like Amazon AWS,Google Cloud,and Microsoft Azure.Given their large scale and recent emergence,they are often at the forefront of electric-ity consumption

58、and efficiency innovations.Hyperscale and co-location centers together account for the lions share of U.S.data center loadabout 60%70%7.143222538|Powering Intelligence May 2024Figure 1.U.S.enterprise data center distribution as of 2022 4,8,9Figure 2.U.S.co-location data center distribution as of 202

59、2 4,8,90258Enterprise Data Centers(number per state)Enterprise Data Centers0163Co-Location Data CentersCo-Location Data Centers(number per state)143222539|Powering Intelligence May 2024Data Centers Primary Electricity-Consuming Hardware and EquipmentThe electricity needs of data centers are determin

60、ed primarily by the three constituent hardware categories.Each categorys proportion of energy consumption can vary depending on the data centers age,configuration,type,and function 10,11,12,13.The three main categories and their energy consumption 2,13,14 are:IT equipment,typically composing 40%50%o

61、f data center energy consumption,encompasses the following foundational hardware units:Servers,which are the workhorses,responsible for data processing and computational tasks Storage systems,which include both traditional hard disk drives(HDDs)and the faster,more energy-efficient solid-state drives

62、(SSDs),crucial for data retention Network infrastructure,which comprises switches,routers,and other components,ensuring seamless data transfer and connectivity Cooling systems,typically composing 30%40%of data center energy consumption,are critical to maintaining an optimal temperature within data c

63、enters to prevent hardware malfunction and ensure longevity.While data centers historically used traditional HVAC,advanced cooling technologies in data centers have transitioned towards systems that are specialized for data center use.Please refer to the section Energy Efficiency and Load Management

64、 below for more details.Auxiliary components,typically composing 10%30%of data center energy consumption,are used for various operational needs and include uninterruptible power supplies,security systems,and lighting.Assessing data center energy efficiency is crucial to gauging how effectively they

65、use electricity.These assessments help to identify trends,drive improvements,and set benchmarks for electricity usage;and play a key role in operational strategy 15,16.Figure 3.U.S.hyperscale data center distribution as of 2022 4,8,9040Hyperscale Data CentersHyperscale Data Centers(number per state)

66、1432225310|Powering Intelligence May 2024AI AND DATA CENTER POWER CONSUMPTION INSIGHTSImmense Volumes of Data are Being Processed DailyData centers worldwide electricity use in 2022 totaled 300 million megawatt-hours(MMWh),or 1.2%of all load,a 45%increase from 2015 17.In the United States in 2023,da

67、ta centers accounted for about 4%of total electricity con-sumption or 150 MMWh,equivalent to the average annual consumption of 14 million households 9,18.Since 2017,annual data volumes have soared,tripling to around 4,750 exabytes(an exabyte being a billion gigabytes)by 2022,showcasing the immense v

68、olume of information being processed and transmitted globally every day 19.In 2022,the daily generation of dataincluding captured,copied,or consumedreached approximately 13 exabytes,a surge partly attributable to the burgeon-ing impact of AI models 17.Concurrently,in 2022,global data transmission ne

69、twork energy use was reported to be around 260360 MMWh,roughly equal to data center power use 17,20.Figure 4 illustrates the dramatic rise in global consumer IP traffic.Data centers are facing a significant challenge with internet traffic growing nearly 12-fold in the past decade,a trend paralleled

70、by increasing AI-related workload demands 19.The historical precedent is showcased in Figure 5,which contrasts the U.S.data storage supply versus estimated demand from 2009 to 2020,underscoring a growing deficit and the need to address these trends 22.Despite the immense growth in network traffic an

71、d data generation,load growth has been much slower due to ef-ficiency gains and consolidation.Figure 4.Trends in global consumer IP traffic,20172022 21Figure 5.U.S.data storage supply vs.demand,20092020 22,230100200300400Data volume in exabytes per month201720182019202020212022Data volume of global

72、consumer IP traffic from 2017 to 2022122156201254319396035,00030,00025,00020,00015,00010,0005,00040,00045,000DemandSupplyStorage Capacity in Exabytes2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 20201432225311|Powering Intelligence May 2024History of Energy Efficiency in the Data Center Ind

73、ustryOver the last 25 years,U.S.data center load growth,as shown in Figure 6,has experienced three phases:1.Energy consumption grew in the early 2000s driven by the rapid expansion of internet infrastructure and the dot-com boom 24.2.From 20102020,electricity consumption stabilized as data center ex

74、pansion was offset by equally rapid im-provements in energy efficiency achieved both through improvements at individual facilities and through the transition from small data centers to more efficient cloud facilities 25,26.3.Recent load growth in data centers is driven mainly by the expanding demand

75、 for cloud services,big data ana-lytics,and AI technologieswhich require significant computational resourcesand a slowing of efficiency gains 27.Efficiency gains in individual data centers have been led by advancements in server efficiency,which have been significant,leading to reduced power consump

76、tion per unit of computing power 28.Power and cooling equipment,required to operate the IT components,has also improved its efficiency.Power Usage Effectiveness(PUE)and Data Center Infrastructure Efficiency(DCIE),key efficiency metrics in the data center industry,are defined in the box on the next p

77、age.Figure 7 shows the U.S.PUE declined from 2007 to 2022,illustrating notable efficiency gains in Figure 7.Average annual PUE in data centers,20072023 26,29Figure 6.U.S.data center load growth from 2000 to 2023.The graphs light blue area indicates the uncertainty range based on two datasets estimat

78、ing recent-to-current data center loads 4,8,9MaxMin200150100500Electricity Consumption(TWh/y)2000200420082016201220202024Average Annual PUE20072.501.981.651.581.581.671.591.571.55201120142018201920202021202220231432225312|Powering Intelligence May 2024non-compute power demands 26,29.The recent stabi

79、lization at an average PUE of 1.6 suggests slowed progress in energy-saving strategies amidst the recent,rapid buildout 29.With increased cooling needs of GPUs,advanced technologies(as discussed later in the report)could restart the downward drive in cooling and ancillary equipment power needs,with

80、projected PUEs in advanced facilities potentially approaching 1.2.Uneven Geographic Distribution Creates Imbalance in Data Center LoadThe geographic distribution of data centers is notably uneven,creating economic opportunity but also localized grid stress.For example,in 2022,data centers accounted

81、for 1.2%of worldwide electricity,but 20%of electricity con-sumption in Ireland 30.Similarly,the United States shows uneven growth in data center investments,reflecting a diverse landscape of regional opportunities and challenges.Data centers consume more power in Virginia,for instance,than in any ot

82、her state.9,17Fifteen states accounted in 2023 for 80%of the national data center load.Ranked from highest to lowest in estimat-ed load,each presents both opportunities and challenges as shown in Table 1.Table 1.Opportunities and challenges for states ranked in the Top 15 for data center growth 29,3

83、1,32,33,34,35,36,37STATEOPPORTUNITIESCHALLENGESVirginiaUnparalleled network infrastructure;proximity to federal government agenciesCommunity pushback;regulatory scrutiny,particularly concerning environmental impact;transmissionTexasBusiness-friendly;ample land availabilityElectric grid reliability a

84、nd pace of expansionCaliforniaRobust technological ecosystemHigh real estate and power costs;stringent environmental regulationsIllinoisStrategic location;significant tax incentives;nuclear generation and increasing renewable energy investments to address sustainabilityTransmission constraints and r

85、apidity of developmentOregonLow electricity rates,low carbon emissions,moderate climate,tax incentives,and skilled workforceComplex environmental regulations;demand for green energy solutions,and pace of growthArizonaSolar electricity,low risk of natural disasters;recent market growthWater scarcity;

86、need for sustainable cooling solutionsIowaLow electric rates;renewable energy availabilityGeographic limitation(distant from major U.S.data hubs)GeorgiaAvailability of land and power;friendly business environmentBalancing rapid expansion with local resource impactsWashingtonAbundant renewable energy

87、 resourcesHigh energy costs;strict regulatory measuresPennsylvaniaStrategic location near major East Coast markets;relatively low energy costsAging infrastructureNew YorkHub for financial services;high connectivity demandSpace limitations;high energy costsNew JerseyClose to major metropolitan areas;

88、robust fiber-optic infrastructure;transmission capacity from recent build outHigh property and energy costsNebraskaLow energy costs;generous tax incentivesRemote location might limit connectivity optionsNorth DakotaSignificant tax incentives;low cost of operationsLimited connectivity;need for more r

89、obust infrastructureNevadaTax abatements;low electricity pricesWater scarcity;need for sustainable cooling solutionsDEFINITIONS OF KEY EFFICIENCY METRICS FOR DATA CENTERS Power Usage Effectiveness(PUE)A metric that quantifies a data centers energy efficiency by dividing the total energy usage by the

90、 energy consumed by the IT equipment alone.A lower PUE indicates higher efficiency,with 1.0 implying that all energy use is for computation.Data Center Infrastructure Efficiency(DCIE)A measure of a data centers energy efficiency calculated as the percentage of energy used directly by IT equipment ou

91、t of the total energy consumption.Higher DCIE values signify greater efficiency in non-computational functions.1432225313|Powering Intelligence May 2024Table 2 presents estimates of data center consumption in 2023,2030,and the projected consumption as a percent-age of state electricity demand(%EC)fo

92、r the 15 states.For detailed graphs of each states projections as well as a table showing 44 states that are pertinent to the U.S.data center market,see Appendix A.Table 2.Current and projected load growth in Top 15 states 4,8,9FORECASTED SCENARIOS:PROJECTIONS OF DATA CENTER ELECTRICITY CONSUMPTION

93、IN TOP 15 STATES(20232030)STATE2023 LoadLow-growth cenario(3.71%)Moderate-growth cenario(5%)High-growth cenario(10%)Higher-growth cenario(15%)MWh/y%of Total State Electricity Consumed(%EC)MWh/y%of Total State Electricity Consumed(%EC)MWh/y%of Total State Electricity Consumed(%EC)MWh/y%of Total State

94、 Electricity Consumed(%EC)MWh/y%of Total State Electricity Consumed(%EC)Virginia33,851,122 25.59%43,683,508 29.28%47,631,928 31.10%65,966,260 38.47%89,880,357 46.00%Texas21,813,159 4.59%28,149,002 5.47%30,693,306 5.94%42,507,676 8.04%57,917,564 10.64%California9,331,619 3.70%12,042,078 4.43%13,130,5

95、25 4.81%18,184,686 6.54%24,777,000 8.70%Illinois7,450,176 5.48%9,614,151 6.53%10,483,145 7.08%14,518,285 9.54%19,781,455 12.56%Oregon6,413,663 11.39%8,276,574 13.39%9,024,668 14.43%12,498,415 18.93%17,029,342 24.14%Arizona6,253,268 7.43%8,069,590 8.81%8,798,975 9.53%12,185,850 12.73%16,603,465 16.58

96、%Iowa4,828,44011.43%6,230,907 13.44%6,794,100 14.48%9,409,26318.99%12,820,31024.21%Georgia6,175,391 4.26%7,969,093 5.08%8,689,396 5.51%12,034,090 7.48%16,396,690 9.92%Washington5,171,612 5.69%6,673,757 6.77%7,276,977 7.34%10,078,009 9.88%13,731,490 13.00%Pennsylvania3,594,0383.16%4,637,9613.78%5,057

97、,172 4.11%7,003,7635.61%9,542,7687.49%New York4,067,385 2.84%5,248,796 3.40%5,723,219 3.69%7,926,182 5.05%10,799,583 6.75%New Jersey 3,723,199 5.42%4,804,638 6.46%5,238,9147.00%7,255,4619.44%9,885,711 12.44%Nebraska2,896,29511.70%3,737,552 13.75%4,075,37814.81%5,644,05919.41%7,690,14524.71%North Dak

98、ota2,975,81515.42%3,840,169 18.00%4,187,271 19.31%5,799,02224.89%7,901,28431.11%Nevada3,416,7078.69%4,409,12210.28%4,807,64911.10%6,658,19514.75%9,071,92419.07%*The four load growth projection scenarios reflect national-level estimates of data center growth applied to state-level estimates of curren

99、t demand.This analytical approach is explained in more detail in the section Forecasting Data Center Load Growth to 2030.1432225314|Powering Intelligence May 2024The map in Figure 8 depicts the projected data center share of state electricity demand in 2030,calculated by applying the annual U.S.data

100、 center growth rates(averaged across the four scenarios)to project state-level data center loads and assuming other loads grow at 1%annually.(The sce-narios are explained in the section Forecasting Data Center Load Growth to 2030.)The potential for a rapidly rising share of data center power demand

101、in many states accentu-ates the need for customized energy strategies that align with the specific demands and infrastructure capabilities of each states grid.State-level projections also underscore the critical need for innovation in energy management and the optimization of localized infrastructur

102、e to accommodate the rising energy demands associated with expanding data center workloads.AI Implications for Power ConsumptionIn the latter half of the 20th century,AI applications typi-cally involved rule-based strategies and small machine-learning models that used very little electricity.However

103、,as the 21st century unfolded,AI systems witnessed exponen-tial growth in their complexity and computational require-ments 38,39.On a global level,the United States has been leading in the development of prominent AI systems,with the creation of 16 such systems since 2022,compared to the United King

104、doms eight and Chinas three 39.Key AI-related technological drivers contributing to escalat-ing data center electricity demands include:1 The exponential growth of data generation:The dramatic rise in global consumer IP traffic represents a reflection of the“big data”wave,part of which has resulted

105、from feeding AI models with diverse and large datasets 9,10,40,41.The surge in data avail-ability not only fuels the sophistication and accuracy of AI algorithms but also underscores the symbiotic relationship between increasing internet usage and AI advancement.Of course,this has required expanded

106、storage,increased processing capabilities,and escalat-ing electricity demands 21.The increasing complexity of AI models:Initially consti-tuted as rule-based entities functioning through coded 1 While cryptocurrency mining,with its distinct computational pro-cesses and energy patterns for blockchain

107、transaction verification and cryptocurrency generation,also impacts energy loads,it is excluded from this study to maintain focus on traditional data center opera-tions and AI-driven computations.In 2022,global crypto mining was estimated to have consumed around 110 million MWh,accounting for 0.4%of

108、 annual global electricity demand,around one-third the usage of traditional data centers 17.A separate assessment is warranted to understand the potential power needs and flexibility of cryptocurrency power demands.Figure 8.2030 projected data center share of electricity consumption(assumes average

109、of the four growth scenarios and that non-data center loads grow at 1%annually)4,8,905%510%1015%1520%20+%2030 Data Center%of State Electricity Consumption1432225315|Powering Intelligence May 2024instructions,AI models have undergone a monumental transformation,becoming increasingly complex and capab

110、le over time 42,in turn increasing their compu-tational demands.As an illustration of the staggering increase in computation demand,note that in 1957,the Perceptron Mark I,the first real-world implementa-tion of a one-layer neural network that could classify images,utilized 695,000 floating-point op

111、erations per second(FLOPS)an assessment of AI complexity and computation intensity.In 2020,however,GPT-3 re-quired a staggering 3.14x1023 FLOPS,an increase of 18 orders of magnitude,and at present each subsequent AI model is requiring even greater amounts.The continuous operational demands of a digi

112、tal ecosystem:In the modern era,data centers function ceaselessly to uphold the demands of a globalized soci-ety that thrives on connectivity.Data centers facilitate uninterrupted services,ensuring 24/7 availability in various sectors including business,e-commerce,and entertainment.Maintaining const

113、ant uptime requires robust backup power solutions.Energy contributions of AI annual workloads are catego-rized into three major areas 7,18,39,43,44:Model development(10%of the energy footprint):Models are developed and fine-tuned before training.Model training(30%of the energy footprint):Algo-rithms

114、 learn by processing a vast array of data to make predictions or decisions without exact input-response relations preprogrammed,which requires substantial computation efforts and high energy expenditure for extended periods.Use/inference(60%of the energy footprint):Includes the deployment and utiliz

115、ation of developed AI models in real-world applications and requires computational resources for interpreting new data and generating outcomes or predictions based on pre-trained models.For detailed information on AI model types,specific models and their descriptions,and the electricity consumption

116、of each,see Appendix B.Chat GPT and Other Large Language Models(LLMs)Over the last year,the surge in popularity of generative AI sparked by the public release of Open AIs ChatGPT has created new concerns about AIs potential impact on future computing energy needs.Figure 9 shows the increase in web t

117、rafficstarting from zerofor prominent generative AI platforms including ChatGPT,which is illustrated by the dark blue line 45.ChatGPT garnered 100 million global users in only two months,which was rapidly followed by tech giants like Microsoft,Alphabet,Meta,and Bing launching their own large-languag

118、e model chatbots.From a power usage per-spective,these LLMs create a new frontier with ultimate impact to be determined,in part,by how widely the 5.3 billion internet users adopt the new features being rolled out.46,47.For example,Google plans to implement LLMs to boost its search engines ability to

119、 recognize and respond to user queries in a more conversational and natural style 48.At Figure 9.U.S.web traffic trends to AI platforms,20222023 4501020304050ChatGPT 4.0Character AIGoogle BardMillions weekly web visits,USNov18Dec1620222023Jan13Feb10Mar10Apr7May5Jun2Jun30Jul28Aug25Sep22Oct20143222531

120、6|Powering Intelligence May 20242.9 watt-hours(Wh)per ChatGPT request,AI queries are estimated to require 10 times the electricity of traditional Google queries,which use about 0.3 Wh each 47.Imple-menting LLMs in every Google search could necessitate 80 gigawatt-hours(GWh)daily or 29.2 terawatt-hou

121、rs(TWh)yearly electricity consumption,according to SemiAnalysis 34.New Street Researchs similar analysis suggests the need for around 400,000 servers,consuming 62.4 GWh daily or 22.8 TWh yearly 47.As shown in Figure 10,the BLOOM models electricity usage averages 3.96 Wh per request,while ChatGPTs is

122、 slightly lower at 2.9 Wh per request;however,if Google integrated similar AI into its searches,the electricity per search could increase to be-tween 6.98.9 Wh 47.The explosive growth in investments aimed at building and deploying new AI capabilities are raising concerns over the overall electricity

123、 consumption and environmental impact of AI and data centers and the ability of the United States to maintain its leadership position.Figure 10.Electricity consumption per request 470246810Wh per requestGooglesearchChatGPTBLOOMAI-poweredGoogle search(New StateResearch)AI-poweredGoogle search(SemiAna

124、lysis)1432225317|Powering Intelligence May 2024FORECASTING DATA CENTER LOAD GROWTH TO 2030Four Scenarios Based on Historical Data,Expert Insights,and Current TrendsDrawing on public information about existing data centers,public estimates of industry growth,and recent electricity demand forecasts by

125、 industry experts,EPRI prepared four projectionsusing low(3.7%),moderate(5%),high(10%),and higher(15%)growth scenarios described in Table 3 belowof potential electricity consumption in U.S.data centers from 2023 to 2030.See Figure 11 for a graph of the projections.These projections are based on a bo

126、unding analysis of various data sources surveyed as of November 2023 1,2,4,8,14.The analysis reflects historical trends for the AI industry,internet traffic,demand for storage,coupled with the computational intensity and prevalence of AI models.All of these factors are uncertain,including the develo

127、pment of business models and updates for LLMs,rate of increase in mature applications,and efficiency gains in computational and non-computational aspects of data centers.The graphs blue line depicts average historical data center electricity consumption.The light blue area indicates the uncertainty

128、in recent historical projections of data center power use,and the colored swaths show the four projec-tion scenarios 4,8.Under the 15%higher growth scenario,EPRIs projections show data center electricity usage rising to an average of 403.9 TWh/year.Under the 10%high growth scenario,data center energ

129、y usage rises to a mid-range of 296.4 TWh/yr.Using the moderate growth 5%scenario,the projec-tion predicts a mid-range of 214.0 TWh/yr.Under the 3.7%low growth scenario,the graph shows the projection at a mid-range of 196.3 TWh/yr.The mid-range estimates of data centers share of total U.S.electricit

130、y consumption in 20309.1%,6.8%,5.0%,and 4.6%assume that other loads grow at 1%annually.An examination of regional variations is found in Appendix A.Figure 11.Projections of potential power consumption in U.S.data centers scenarios,20232030 1,2,4,8,146005004003002001000Electricity Consumption(TWh/y)2

131、000200520102015202020252030MaxMinAverage historical data4.6%5.0%6.8%9.1%3.7%5%10%15%Low growthModerate growthHigh growthHigher growthSCENARIOANNUALGROWTH RATE%OF 2030 ELECTRICITYCONSUMPTION1432225318|Powering Intelligence May 2024Each growth scenarios characteristics are described in Table 3.The loa

132、d projections combine estimates of todays data center power usage with assessments of potential future technological advances and computational demands.It is essential to understand that while these scenarios are based upon the latest available data and subject-matter expert(SME)insights,the factors

133、 affecting themsuch as consumer demand,technological advancements,opera-tional efficiencies,and evolving industry standardsare changing almost daily.ENERGY EFFICIENCY,LOAD MANAGEMENT AND CLEAN ELECTRICITY SUPPLYWith the escalating demands of AI and data center opera-tions,there is a critical need fo

134、r new,innovative strategies that leverage advances in hardware,system monitoring,computational algorithms,clean electricity procurement,and operational flexibility 49.Key considerations include:Energy efficiency:Adopt advanced cooling solutions,power management systems,and leverage efficiency advanc

135、es in computational and supporting hardware to reduce overall electricity consumption.Scalability:Implement modular designs and virtual-ization techniques to ensure that infrastructure can handle future demands without disproportionate increases in energy use.Carbon-free energy(CFE)use:Transition to

136、 carbon-free electricity sources for data center operations and low-carbon technologies for backup power to support data center expansion without increasing carbon emissions and make energy costs more certain.Monitoring and analytics:Utilize real-time monitoring tools to track electricity consumptio

137、n,detect inefficien-cies,and optimize operations.Research and development(R&D):Invest in innovations that drive both performance and sustainability,such as green energy sources or AI-driven optimization of work load to meet latency,grid,environmental,and other objectives.The remainder of this sectio

138、n discusses in more detail strat-egies that cover energy-efficient algorithms,hardware,and cooling technologies;scalability and renewable energy use;and monitoring,analytics,and R&D.Energy-Efficient Training AlgorithmsThe initial phases of AI algorithm development have been heavily focused on enhanc

139、ing accuracy and augmenting performance capabilities.However,as the performance of the algorithms has increased and recognizing the expo-nential growth in computational demands,the paradigm is beginning to shift to also value the efficiency of model development.Recent studies document applications w

140、here a slight compromise on model accuracy has yielded sub-stantial reductions in electricity consumption 7,18,43,50.The techniques utilized include:Pruning:This technique aims to reduce or eliminate unnecessary elements in neural networks,thereby maintaining robust performance while reducing com-pu

141、tational complexity 43,51.Quantization:This method reduces the numerical preci-sion of computations,effectively conserving electricity without compromising significantly on accuracy 14,51.Knowledge distillation:This approach involves devel-oping a smaller,more manageable model that mirrors the funct

142、ionalities of a larger,more intricate structure,reducing computational requirements 14,51.Table 3.Forecasted load projections:Parameters of power consumption in each of the four U.S.data center scenarios,2022-2030 4,8,9COMPOSITION OF GROWTH SCENARIOS(20232030)GROWTH SCENARIOCAGR(%)AVERAGE 2023 DATA

143、CENTER LOAD(MWH)AVERAGE PROJECTED LOAD,2030(MWH)CHANGE IN GROWTH()Higher Growth15%152,120,846403,906,136166%High Growth10%152,120,846296,440,49395%Moderate Growth5%152,120,846214,049,30641%Low Growth3.7%152,120,846196,305,81829%1432225319|Powering Intelligence May 2024Energy-Efficient HardwareComput

144、ational hardware is becoming more efficient,venturing beyond general-purpose central processing units(CPUs)to embrace an array of specialized hardware.These hardware variants are customized for specific tasks,stream-lining power usage,and enhancing overall efficiency.This specialized hardware includ

145、es:Tensor processing units(TPUs):Specifically designed to expedite machine learning(ML)tasks,these units provide pronounced performance and energy efficiency enhancements 11,18.For example,Googles Cloud TPUv4 showed not only a 10-times leap forward in ML system performance over TPUv3,but it also boo

146、sted energy efficiency by 23%compared to contemporary ML data structures and algorithms 52.Field-programmable gate arrays(FPGAs):Recognized for their versatility as non-hard etched processors,FPGAs can be reprogrammed for specific tasks,provid-ing improved performance and lower per-unit energy consu

147、mption 11,28.Though savings are task-depen-dent,FPGAs have shown reductions in memory and bandwidth usage as much as 75%when compared to traditional CPUs and graphics processing units(GPUs)28,53,54.Power capping:Some processing chips,such as GPUs,can operate at reduced power levels.For example they

148、can reduce direct power consumption by 10%while also reducing cooling needs.Energy-Efficient Cooling TechnologiesHeat is a byproduct of computation,and traditional cooling methods are energy-intensive,composing around 35%of data center electricity use.However,innovative solutions are emerging,some o

149、f which include:Liquid cooling:Utilizing liquids to absorb and dissipate heat can use less electricity than traditional air-cooling systems 18,50.A recent study,which examined the shift from 100%air cooling to a combination of 25%air cooling and 75%liquid cooling,highlighted the efficien-cy gainslea

150、ding to a notable decrease in PUEfrom transitioning to hybrid cooling systems in data centers.The study observed a 27%reduction in facility power consumption and a 15.5%decrease in overall energy usage across the data center site 55.Table 4 shows an overview of various innovative cooling ttechnologi

151、es currently being adopted or considered in data centers,highlighting vendor-reported technology-readiness level(TRL)and energy-saving estimates 56,57,58,59,60,61.Economizer use:An economizer can evaluate outside temperature and humidity,and use exterior air to help cool data center infrastructure w

152、hen appropriate,minimizing reliance on mechanical cooling methods and leading to significant electricity savings 18,62.A 2015 study found that air-side economizers yielded cooling coil load savings of 7699%in comparison to conventional cooling systems in data centers;and the total cooling energy sav

153、ings of the economizers ranged from 47.5%67.2%62.Table 4.Emerging cooling technologies with vendor-reported TRLs and energy savings 56,57,58,59,60,61EMERGING COOLING TECHNOLOGIESTECHNOLOGYTECHNOLOGY-READINESS LEVEL(TRL)CLAIMED EFFICIENCY DIFFERENTIAL(%)Air-Assisted Liquid Cooling9This technology off

154、ers up to a 50%reduction in energy usage compared to traditional air cooling,with the potential to reach a PUE of less than 1.1 56.Immersion Cooling 8Immersion cooling promises substantial energy savings from 5095%compared to traditional air-cooling methods 57,58.Microconvective Liquid Cooling6This

155、emerging technology proposes an 18%energy saving and a PUE of 1.02,alongside a 90%reduction in water usage compared to other liquid systems,indicating its potential for more sustainable operation 59.Radiative Cooling6This solution offers 5070%energy savings,with the benefits of zero water use and lo

156、w maintenance 60.Two-Phase Liquid Immersion Cooling7This technology claims a 41%energy saving compared to air cooling,noting its water conservation and space-saving benefits 61.1432225320|Powering Intelligence May 2024 Heat reuse:Heat generated by computation can be used for various applications suc

157、h as heating adjacent buildings,particularly in cold climates,thereby reduc-ing overall energy usage 18,42,63.Since 2016,Ama-zons 1.1 million-square-foot Doppler building has been estimated to recover 3200 MWh of excess heat from a nearby data center;this is projected to continue over the next 25 ye

158、ars.This heat,which would otherwise have been wasted and would have required cooling equipment,is redirected through the districts energy system,demonstrating an energy-efficient approach to energy reutilization 25.Scalable Clean Energy UseAs digital services proliferate and demand for computa-tiona

159、l power intensifies,scalable clean energy supplies are important to avoid increases in greenhouse gas emissions 64.Corporate commitments to acquire carbon-free elec-tricity on an annual or hourly-matched basis are emerging and can play a significant role in reducing data center emis-sions impacts.Th

160、ese include:Clean electricity procurement from the grid and clean onsite generation:Data center owners have been instrumental in driving the corporate shift towards con-tracting for renewable energy to provide their power needs.In 2021,Apple,Google,Meta,and Microsoft matched their operational electr

161、icity consumption,predominantly from data centers,on an annual basis with the purchase or generation of renewable electric-ity2800 MWh,18,300 MWh,9400 MWh,and 13,000 MWh respectively 55,56,57,58.Meanwhile,Ama-zons operations consumed 30,900 MWh,85%of which was matched on an annual basis by generatio

162、n from re-newable sources,and the company aims to reach 100%renewable energy by 2025 25,16,41.Moreover,a growing number of organizations are working towards 24/7 CFE,which entails matching their electricity de-mand with carbon-free sources in the same region on an hourly basis.This hourly matching w

163、ill require flex-ible technologies such as batteries that can shift solar or wind output to times when they are needed as well as firm clean capacity such as nuclear,fossil plants with carbon capture and storage,or geothermal,that typi-cally operate around the clock.Spurring deployment of flexible a

164、nd clean firm assets can help speed the path to a net-zero power sector 42,44.Cleaner onsite backup power systems:Backup power systems at most existing data centers typically operate for less than 100 hours annually when the grid or pri-mary power supply are unavailable.Accordingly,they constitute o

165、nly a small portion of a data centers envi-ronmental footprint.Shifting from the most common backup technology,diesel generators,to lower-emitting alternatives,like battery energy storage systems(BESS)or cleaner fuelssuch as renewable natural gas,bio-diesel,or clean hydrogen or ammonia,especially wh

166、en the latter are integrated with fuel cellscan reduce backup GHG emissions and,in some instances,allow more frequent operation of these resources,creat-ing the potential for them to serve as a grid resource when/if needed 9,42.Clean onsite or nearby technologies such as nuclear generation or renewa

167、ble generation coupled with long-duration energy storage that can match the growing size of data centers:With currently proposed data cen-ters reaching 1 GW or more at a single site,the scale of power demand is escalating rapidly.In the near term,uprating,relicensing,or restarting existing nuclear p

168、lants near data centers could provide one solution.Amazons purchase of a data center in Pennsylvania co-located with a Talen nuclear power plant provides one example of utilizing existing nuclear.Looking forward,small modular reactors(SMRs)offer a scalable power solution that can grow with the deman

169、ds of a data cen-ter.Companies such as NuScale are exploring scalable capacities of 250600MW for SMRs 9,42.Standard Power has chosen NuScales SMR technology to power two facilities it plans to develop,one in Ohio and the other in Pennsylvania 69.However,while SMRs might supplement power supplies in

170、the future,their waste output,operational risks,and regulatory challenges call for a comprehensive assessment of benefits against potential environmental concerns.Monitoring and AnalyticsAdvances in monitoring and analytics of power consump-tion play a crucial role in realizing operational savings i

171、n data centers.These processes enable precise tracking of energy usage,identification of inefficiencies,and implemen-tation of advanced technologies,thus driving cost reduction and enhancing overall efficiency:Efficient server management:Traditionally,data centers have grappled with up to 30%server

172、underutilization,where servers consume energy but dont fully utilize their computational capabilities.However,with the adoption of innovations like advanced scheduling and 1432225321|Powering Intelligence May 2024dynamic resource allocation,some companies are aiming to reduce underutilization rates

173、to below 10%within the next five years 18,40,63,70.In addition,the implementation of virtualization and containeriza-tion can enhance server efficiency significantly,po-tentially increasing server capacity utilization by 45%by having a single physical server handle more work-loads through virtual or

174、 containerized environments.If successful,this is estimated to reduce the number of physical servers needed,leading to about 20%less energy consumption per unit of computation over the next decade 15,33,42.Flexible computation strategies:Optimizing data center computation and geographic location to

175、respond to electricity supply conditions,electricity carbon inten-sity,and other factors in addition to minimizing latency enables data centers to actively adjust their electricity consumption 71.For example,some could achieve significant cost savingsas much as 15%by optimiz-ing computation to capit

176、alize on lower electric rates during off-peak hours,reducing strain on the grid dur-ing high-demand periods 38,72.With technological and regulatory advances,these strategies could evolve to incorporate real-time energy market dynamics en-abling data centers to not only adjust their operations based

177、on grid demands but also actively participate in energy markets to optimize their benefits and support grid stability.Reducing Data Centers Environmental FootprintThe previous sections focus on actions that data center owners and operators are actively pursuing to diminish their carbon footprint,foc

178、using primarily on onsite direct emissions such as from onsite generation(Scope 1 emis-sions)and emissions associated with the purchase of electricity(Scope 2 emissions)64.These strategies involve reducing their electricity needs through the adoption of ad-vanced computation,cooling,and operational

179、technologies,shifting toward cleaner onsite backup power,and moving towards various strategies for matching their hourly loads with carbon-free electricity 73.Several of the hyperscale companies have fully matched their annual power purchas-es with carbon-free electricity on an annual basis and are

180、moving forward on hourly matching.Progress is slower on shifting to cleaner backup power(although this,as noted earlier,represents only a small fraction of their environ-mental footprint).In recent years,some companies have taken the additional step of quantifying and setting reduction targets for t

181、heir(Scope 3)indirect emissions,which include emissions asso-ciated with supply chains and end-user services 7,44,66.Key actions include sourcing materials from environmen-tally responsible vendors,minimizing the carbon footprint associated with transportation and logistics,and ensuring that the lif

182、ecycle of data center components is managed sustainably,from manufacturing to end-of-life disposal and recycling 42,74,75.ACTIONS TO SUPPORT RAPID DATA CENTER EXPANSIONData centers are one of the fastest growing industries worldwide.These facilitiesand advanced cloud comput-ing and AI technologies t

183、hat are proliferating and driving further growthrepresent large point loads and are at the leading edge of an anticipated global rise in electricity demand driven by efficient electrification and production of low-carbon fuels.In the United States,data center power demand growth,coupled with increas

184、ing electricity demands from EVs,heat pumps,electrification in industry,and the onshoring of manufacturing incentivized by the CHIPS Act,Inflation Reduction Act(IRA),and Infrastructure Investment and Jobs Act(IIJA),is placing both immediate and sustained pressure on the electric grid to accommodate

185、new loads.Clusters of new,large point loads create several challenges.Data centers speed from breaking ground to operationoften within two or three yearsrequirements for highly reliable power,and requests for power generated by new,non-emitting generation sources can create local and re-gional elect

186、ric supply challenges and test the ability of elec-tric companies to keep pace.The most serious challenges to data center expansion are local and result from the scale of the centers themselves and mismatches in infrastructure timing.EPRI highlights three essential strategies to support rapid data c

187、enter expansion.These strategies,each of which is explained below,emphasize increased collaboration be-tween data center developers and electric companies and are:Improve data center operational efficiency and flex-ibility Increase collaboration through a shared energy econo-my model for sustainable

188、 data centers1432225322|Powering Intelligence May 2024 Better anticipate future point load growth through improved forecasting and modelingImprove Data Center Operational Efficiency and FlexibilityOver the past decade,economy-wide electricity demand in the United States has remained relatively flat

189、in large part due to enhanced energy efficiency,which has offset poten-tial increases driven by economic expansion and population growth.Specific to data centers,power demands from rapid expansion in computation,communication,and data stor-age were largely offset by efficiency gains for over a decad

190、e.This is largely due to technological advancements in com-putation,improved cooling systems,sophisticated energy management strategies,and the replacement of many small data centers with more efficient cloud data centers.However,since around 2018,efficiency gains have slowed,data center expansion a

191、ccelerated(in part due to lifestyle changes caused by the pandemic),and AI has proliferated,leading to an increase in data center power consumption.Meeting the increasing electricity demands of AI and data centers while limiting the growth of CO2 emissions neces-sitates a comprehensive strategy that

192、 intertwines techno-logical advancements that improve efficiency with power purchase and production strategies that favor low-carbon resources and that increase both temporal and spatial flex-ibility to link intense operation periods to the availability of low-cost,low-carbon generation.Computationa

193、l efficiency gains require investing in the next generation of energy-efficient processors and server archi-tectures and enhancing AI training algorithms for greater computational efficiency.From an architectural viewpoint,virtualization stands out,with its capability to run multiple virtual machine

194、s on one physical server,potentially cut-ting hardware needs by 3040%with consequent electric-ity savings 9,75.Implementations like software-defined infrastructure(SDI)offer dynamic resource allocation in real time,potentially increasing allocation efficiency by 30%,potentially increasing spatial fl

195、exibility in computation loads.Hybrid cloud solutions provide a balance between on-premises infrastructure and shared cloud services,potentially providing locational flexibility by reducing onsite requirements by 25%during peak periods.In addition,continued gains in data center infrastructure effici

196、ency can be achieved through more effective cooling technologies,adopting energy management systems that leverage AI for optimized power usage,and setting stringent industry targets for energy consumption.Continuous moni-toring and analytics can help data centers better anticipate and react to dynam

197、ic energy needs,ensuring optimal operational efficiency and rapid adaptability 42,50,78.Embedding real-time monitoring tools within AI and data center ecosystems can facilitate immediate insights into fluctuations in electricity usage.Pilot projects to explore and validate novel energy conservation

198、methods,which document and disseminate findings broadly,can accelerate adoption of proven sustainable strategies 10,70.Increase Collaboration through a Shared Energy Economy Model for Sustainable Data CentersElectric companies are challenged as they must meet the increasing and uncertain load from d

199、ata centers while also ensuring reliability,affordability,and sustainability for all customers.Developing a deeper understanding of data cen-ter power needs,timing,and potential flexibilitieswhile assessing how they match available electric supplies and delivery constraintscan create workable soluti

200、ons for all.EPRI,in collaboration with major data center builders/oper-ators/owners and the electric companies that power these facilities,is exploring sustainable approaches to powering the growing wave of AI data centers.Enabled by technol-ogy and supporting policies,data center backup generators,

201、powered by clean fuels,could support a more reliable grid while reducing the cost of data center operation.Shifting the data center-grid relationship from the current“passive load”model to a collaborative“shared energy economy”with grid resources powering data centers and data center backup resource

202、s contributing to grid reliability and flexibil-itycould not only help electric companies contend with the explosive growth of AI but also contribute to affordabil-ity and reliability for all electricity users.This new paradigm of collaboration between data centers and electric companies,which trans

203、forms data centers from passive consumers to active participants in maintaining the grid,is crucial for ensuring electric companies are prepared for the explosive growth of AI.Under this model,data cen-ters move from being a burden on the gridacting as pas-sive loads demanding specific power levels

204、within defined timeframes and at affordable ratesto becoming partners in a sustainable future,serving as a grid reliability resource.The goal is the complete integration of grid and data cen-ter power resources.Clean power generators co-located with data centers act as both grid and data center powe

205、r 1432225323|Powering Intelligence May 2024sources.During grid outages,these resources can seamless-ly form a microgrid to provide uninterrupted power to data centers,eliminating the need and cost of standard diesel backup generators.More research is needed into how data centers and electric grids c

206、an collaborate in a shared energy economy model,as well as the benefits and challenges of doing so.Focusing on U.S.AI training data centers using backup generators pow-ered by clean fuels,EPRI suggests a study of the economic,environmental,social,and technological implications of this shared energy

207、economy model compared to other,more traditional models.The results of this study could provide suggestions and guidelines for data centers and electric grids to adopt and implement the shared energy economy model,or parts of it,in their operations and planning.Better Anticipate Future Point Load Gr

208、owth through Improved Forecasting and ModelingThe lead time for constructing and bringing a large data center online is around two to three years,while adding new electric infrastructure(generation,transmission,sub-stations)can take four or many more years.This highlights the need for better forecas

209、ting and decision tools to antici-pate where and when data center connection requests may appear and characterizing the operational characteristics of that load,especially as the size of interconnection requests grow from hundreds of MW to thousands of MW.In the current environment,electricity compa

210、nies are often receiving multiple requests for the same project from the owner and from developers trying to support the owner.Also,a single data center project may seek interconnec-tion information in multiple locations.And the ramp up to full power demand and operational characteristics on the dat

211、a centers can vary widely,depending upon their func-tion(e.g.,cloud,AI training,AI inference).Therefore,new approaches are needed not only to project where load will grow,but also its operational characteristics and opportuni-ties for flexible operation.EPRIs Load Forecasting Initiative(https:/ in l

212、ate 2023,has research activities underway to help address some of these key uncertainties.1432225324|Powering Intelligence May 2024APPENDIX A:STATE-SPECIFIC SCENARIOSProjected Data Center Load Scenarios for Top 15 StatesFigures A1 through A15 apply the projected U.S.load growth rates under EPRIs hig

213、her-,high-,moderate-,and low-growth scenarios to 2023 estimated state-level data center loads.The figures show projections for the 15 states with the highest data center demands in 2023,comprising around 80%of U.S.data center load in that year.As noted above,the projections utilize the projected nat

214、ional growth rate and do not reflect the deferential regional growth rates implied by Integrated Resource Plan analyses that have emerged recently.Figure A1.Projected electricity consumption in Arizona data centersFigure A3.Projected electricity consumption in Georgia data centersFigure A4.Projected

215、 electricity consumption in Illinois data centersFigure A2.Projected electricity consumption in California data centers51525102030020232030 Projections20222021%of states totalconsumptionElectricity Consumption(TWh/y)Low growthModerate growthHigh growthHigher growth2030 Projections3%5%10%15%16.6%12.7

216、%9.5%8.8%7.4%ArizonaMaxAvg DCMin51525102030020232030 Projections20222021%of states totalconsumptionElectricity Consumption(TWh/y)Low growthModerate growthHigh growthHigher growth2030 Projections3%5%10%15%California8.7%6.5%4.8%4.4%3.7%MaxAvg DCMin51525102030020232030 Projections20222021%of states tot

217、alconsumptionElectricity Consumption(TWh/y)Low growthModerate growthHigh growthHigher growth2030 Projections3%5%10%15%9.9%7.5%5.5%5.1%4.3%GeorgiaMaxAvg DCMin51525102030020232030 Projections20222021%of states totalconsumptionElectricity Consumption(TWh/y)Low growthModerate growthHigh growthHigher gro

218、wth2030 Projections3%5%10%15%Illinois12.6%9.5%7.1%6.5%5.5%MaxAvg DCMin1432225325|Powering Intelligence May 2024Figure A5.Projected electricity consumption in Iowa data centersFigure A6.Projected electricity consumption in Nebraska data centersFigure A7.Projected electricity consumption in Nevada dat

219、a centersFigure A8.Projected electricity consumption in New Jersey data centersFigure A9.Projected electricity consumption in New York data centersFigure A10.Projected electricity consumption in North Dakota data centers51525102030020232030 Projections20222021%of states totalconsumptionElectricity C

220、onsumption(TWh/y)Low growthModerate growthHigh growthHigher growth2030 Projections3%5%10%15%Iowa24.2%19.0%14.5%13.4%11.4%MaxAvg DCMin51525102030020232030 Projections20222021%of states totalconsumptionElectricity Consumption(TWh/y)Low growthModerate growthHigh growthHigher growth2030 Projections3%5%1

221、0%15%Nebraska24.7%19.4%14.8%13.8%11.7%MaxAvg DCMin51525102030020232030 Projections20222021%of states totalconsumptionElectricity Consumption(TWh/y)Low growthModerate growthHigh growthHigher growth2030 Projections3%5%10%15%Nevada19.1%14.8%11.1%10.3%8.7%MaxAvg DCMin51525102030020232030 Projections2022

222、2021%of states totalconsumptionElectricity Consumption(TWh/y)Low growthModerate growthHigh growthHigher growth2030 Projections3%5%10%15%New Jersey12.4%9.4%7.0%6.5%5.4%MaxAvg DCMin51525102030020232030 Projections20222021%of states totalconsumptionElectricity Consumption(TWh/y)Low growthModerate growt

223、hHigh growthHigher growth2030 Projections3%5%10%15%New York6.8%5.1%3.7%3.4%2.8%MaxAvg DCMin51525102030020232030 Projections20222021%of states totalconsumptionElectricity Consumption(TWh/y)Low growthModerate growthHigh growthHigher growth2030 Projections3%5%10%15%North Dakota31.1%24.9%19.3%18.0%15.4%

224、MaxAvg DCMin1432225326|Powering Intelligence May 2024Figure A11.Projected electricity consumption in Oregon data centersFigure A12.Projected electricity consumption in Pennsylvania data centersFigure A13.Projected electricity consumption in Texas data centersFigure A14.Projected electricity consumpt

225、ion in Virginia data centersFigure A15.Projected electricity consumption in Washington data centers51525102030020232030 Projections20222021%of states totalconsumptionElectricity Consumption(TWh/y)Low growthModerate growthHigh growthHigher growth2030 Projections3%5%10%15%Oregon24.1%18.9%14.4%13.4%11.

226、4%MaxAvg DCMin51525102030020232030 Projections20222021%of states totalconsumptionElectricity Consumption(TWh/y)Low growthModerate growthHigh growthHigher growth2030 Projections3%5%10%15%Pennsylvania7.5%5.6%4.1%3.8%3.2%MaxAvg DCMin0204060801001200Electricity Consumption(TWh/y)20232030 Projections2022

227、2021%of states totalconsumptionLow growthModerate growthHigh growthHigher growth2030 Projections3%5%10%15%Texas10.6%8.0%5.9%5.5%4.6%-MaxAvg DCMinElectricity Consumption(TWh/y)020406080100120020232030 Projections20222021%of states totalconsumptionLow growthModerate growthHigh growthHigher growth2030

228、Projections3%5%10%15%Virginia46.0%38.5%31.1%29.3%25.6%MaxAvg DCMin51525102030020232030 Projections20222021%of states totalconsumptionElectricity Consumption(TWh/y)Low growthModerate growthHigh growthHigher growth2030 Projections3%5%10%15%Washington13.0%9.9%7.3%6.8%5.7%MaxAvg DCMin1432225327|Powering

229、 Intelligence May 2024Regional Differences in Data Center Capacities by Metropolitan AreaData center development is heavily clustered in a few counties/cities across the country rather than evenly spread within states,exacerbating power delivery challenges.Figure A16 provides a snapshot for leading

230、metropolitan areas of current data center capacity(measured in MW);additional capacity under development;absorption rates,reflecting the percentage of capacity leased by customers over a specific period of time;and vacancy rates,indicating unutilized space within these data centers.Northern Virginia

231、 is the clear leader in terms of current capacity and current construction.Other regions,such as Dallas-Ft.Worth,Silicon Valley,Chicago,New York Tri-State,and Atlanta,highlight current construction activity that is projected to lead to a 50%or more increase in power demands.29.Figure A16.Data center

232、 development:Key U.S.regions(2022)Under Construction(MW)Current Inventory(MW)Vacancy,%Absorption(MW)Power(MW)NorthernVirginiaDallas-Ft.WorthSiliconValleyChicagoPhoenixNew YorkTri-StateAtlantaCurrent Vacancy(%)05001,0001,5002,0008892,0601%6.1%2.3%6.2%8.5%7.8%3.6%43729939244380380621693424886325441051

233、7818150253332%0%4%6%8%10%1432225328|Powering Intelligence May 2024Projections of Potential Power Consumption for 44 StatesTable A1 presents a detailed view of the energy consumption from data centers in each of the 44 states that had significant data center load in 2023 and contrasts it with project

234、ions for 2030.These projections are categorized into three scenarios:low growth,moderate growth,high growth,and higher growth 1,2,4,8,14.Table A1.Projections to 2030 of potential power consumption for states with significant data center load in 2023 4,8,9FORECASTED SCENARIOS:PROJECTIONS OF POTENTIAL

235、 POWER CONSUMPTION BY STATE(20232030)STATE2023 LoadLow-growth Scenario(3.71%)Moderate-growth Scenario(5%)High-growth Scenario(10%)Higher-growth Scenario(15%)MWh/y%of To-tal State Electric-ity Con-sumed(%EC)MWh/y%of To-tal State Electric-ity Con-sumed(%EC)MWh/y%of To-tal State Electric-ity Con-sumed(

236、%EC)MWh/y%of To-tal State Electric-ity Con-sumed(%EC)MWh/y%of To-tal State Electric-ity Con-sumed(%EC)Alabama1,489,200 1.71%1,921,753 2.05%2,095,454 2.23%2,902,030 3.07%3,954,074 4.13%Arizona6,253,2687.43%8,069,5908.81%8,798,9759.53%12,185,85012.73%16,603,46516.58%California9,331,6193.70%12,042,0784

237、.43%13,130,5254.81%18,184,6866.54%24,777,0008.70%Colorado1,509,6402.66%1,948,1303.18%2,124,2153.46%2,941,8614.73%4,008,3456.34%Connecticut262,8000.95%339,1331.14%369,7861.24%512,1231.71%697,7782.31%Florida1,384,0800.56%1,786,0990.67%1,947,5400.73%2,697,1801.01%3,674,9631.37%Georgia6,175,3914.26%7,96

238、9,0935.08%8,689,3965.51%12,034,0907.48%16,396,6909.92%Hawaii8,7600.10%11,3040.12%12,3260.13%17,0710.18%23,2590.24%Idaho148,9200.57%192,1750.68%209,5450.74%290,2031.03%395,4071.40%Illinois7,450,1765.48%9,614,1516.53%10,483,1457.08%14,518,2859.54%19,781,45512.56%Indiana192,7200.19%248,6970.23%271,1760

239、.25%375,5570.35%511,7040.48%Iowa6,193,32011.43%7,992,23013.44%8,714,62314.48%12,069,02918.99%16,444,29424.21%Kansas8,7600.02%11,3040.03%12,3260.03%17,0710.04%23,2590.05%Kentucky1,620,6002.15%2,091,3192.58%2,280,3472.80%3,158,0913.84%4,302,9625.16%Louisiana78,8400.08%101,7400.10%110,9360.11%153,6370.

240、15%209,3330.20%Maine26,2800.22%33,9130.27%36,9790.29%51,2120.40%69,7780.55%Maryland96,3600.16%124,3490.19%135,5880.21%187,7780.29%255,8520.40%Massachusetts1,062,3692.08%1,370,9442.50%1,494,8602.72%2,070,2573.72%2,820,7665.01%Michigan525,6000.52%678,2660.63%739,5720.68%1,024,2460.95%1,395,5551.28%Min

241、nesota824,3161.24%1,063,7471.49%1,159,8951.62%1,606,3592.23%2,188,6963.01%Missouri972,3601.21%1,254,7911.45%1,368,2081.58%1,894,8552.18%2,581,7772.95%Montana578,1603.71%746,0924.43%813,5294.81%1,126,6706.54%1,535,1118.71%Nebraska3,959,52011.70%5,109,60113.75%5,571,44214.81%7,715,98419.41%10,513,1842

242、4.71%Nevada3,416,7078.69%4,409,12210.28%4,807,64911.10%6,658,19514.75%9,071,92419.07%New Hampshire17,5200.16%22,6090.19%24,6520.21%34,1420.29%46,5190.40%New Jersey4,038,3605.42%5,211,3416.46%5,682,3787.00%7,869,6219.44%10,722,51712.44%New Mexico402,9601.48%520,0041.78%567,0051.94%785,2552.66%1,069,9

243、263.60%New York4,067,3852.84%5,248,7963.40%5,723,2193.69%7,926,1825.05%10,799,5836.75%North Carolina2,672,6761.92%3,448,9812.30%3,760,7242.50%5,208,2893.44%7,096,3994.62%North Dakota3,915,72015.42%5,053,07918.00%5,509,81119.31%7,630,63124.89%10,396,88831.11%Ohio2,363,8861.58%3,050,5001.90%3,326,2252

244、.07%4,606,5452.84%6,276,5103.83%Oklahoma1,226,4001.76%1,582,6202.12%1,725,6682.30%2,389,9073.16%3,256,2964.26%1432225329|Powering Intelligence May 2024FORECASTED SCENARIOS:PROJECTIONS OF POTENTIAL POWER CONSUMPTION BY STATE(20232030)STATE2023 LoadLow-growth Scenario(3.71%)Moderate-growth Scenario(5%

245、)High-growth Scenario(10%)Higher-growth Scenario(15%)MWh/y%of To-tal State Electric-ity Con-sumed(%EC)MWh/y%of To-tal State Electric-ity Con-sumed(%EC)MWh/y%of To-tal State Electric-ity Con-sumed(%EC)MWh/y%of To-tal State Electric-ity Con-sumed(%EC)MWh/y%of To-tal State Electric-ity Con-sumed(%EC)Or

246、egon6,413,66311.39%8,276,57413.39%9,024,66814.43%12,498,41518.93%17,029,34224.14%Pennsylvania4,590,2403.16%5,923,5203.78%6,458,9294.11%8,945,0795.61%12,187,8507.49%Rhode Island17,5200.23%22,6090.28%24,6520.30%34,1420.42%46,5190.57%South Carolina2,023,5602.45%2,611,3232.93%2,847,3523.18%3,943,3464.36

247、%5,372,8885.84%South Dakota70,0800.52%90,4350.63%98,6100.68%136,5660.94%186,0741.28%Tennessee1,327,1401.30%1,712,6211.56%1,867,4191.70%2,586,2202.34%3,523,7773.16%Texas21,813,1594.59%28,149,0025.47%30,693,3065.94%42,507,6768.04%57,917,56410.64%Utah2,562,0377.68%3,306,2069.10%3,605,0449.84%4,992,6861

248、3.13%6,802,63517.08%Virginia33,851,12225.59%43,683,50829.28%47,631,92831.10%65,966,26038.47%89,880,35746.00%Washington5,171,6125.69%6,673,7576.77%7,276,9777.34%10,078,0099.88%13,731,49013.00%Wisconsin148,9200.21%192,1750.26%209,5450.28%290,2030.39%395,4070.53%Wyoming1,857,12011.26%2,396,53813.24%2,6

249、13,15414.27%3,619,00218.73%4,930,96223.90%APPENDIX B:INSIGHTS INTO THE ENERGY USE OF AI MODELSTo better appreciate how AI uses such enormous amounts of electricity,it can be useful to understand more about AI mod-els and how they work.AI models are typically divided into three types:Process automati

250、on and optimization(PAO),which focuses on streamlining and enhancing operations Predictive analytics(PA),which deals with forecasting trends and patterns Natural language processing(NLP),which interprets and generates human languageMoreover,machine learning(ML),a subset of AI,employs statistical met

251、hods to enable machines to improve at tasks with experience.Deep learning(DL),a further subset of ML,involves neural networks with multiple layers that autonomously learn from vast amounts of data.ML and DL have evolved significantly,with industrial applications overtaking academic contributions in

252、recent years.Industrys edge stems from its vast data access,advanced computing capacities,and robust financial backing,positioning it above academia and nonprofit sectors in this subset of the AI domain.MLs broad capabili-ties enable advancements in PAO,PA,and NLP models,while DLs complex neural net

253、works further refine these applications 81,82.Each AI model type has distinct energy implications due to its unique computational requirements.Understanding these distinctions is essential for assessing the broader energy impact of AIs spread.Table B1 provides a comparative analysis of various AI mo

254、dels energy consumption and key characteristics,offering a view of their energy footprints and their computa-tional complexity 44,51,57,63,84,85,86,87,88.1432225330|Powering Intelligence May 2024Table B1.Comparative analysis of AI model load consumption and characteristics 44,51,57,63,84,85,86,87,88

255、LOAD CONSUMPTION:BY SPECIFIC AI MODELMODEL NAMEAI TYPEYEARTRAINING(DAYS)CONSUMPTION(MWH)MODEL DESCRIPTIONT5PA20192085.7A versatile model trained to convert text inputs into text outputs,suitable for various tasks like translation and summarization.MeenaNLP201930232A chatbot model developed by Google

256、 designed to engage in conversations and understand context more naturally.Evolved TransformerPAO201977.5A machine learning model designed using neural architecture search for improved performance on tasks.Switch TransformerPAO202027179A variant of the Transformer model designed to handle a large nu

257、mber of parameters more efficiently by dynamically routing activations to a subset of experts.GShard-600BPAO2020324.1Googles model optimized for large-scale multitask training,aiding in handling vast amounts of parameters.ChatGPT-3NLP2021341,287A state-of-the-art language model by OpenAI known for g

258、enerating coherent and contextually relevant sentences over long passages.BERTPA202162.8A model that understands the context of words in a sentence by analyzing them in both directions(left-to-right and right-to-left),widely used in sentiment analysis and other prediction tasks.GopherNLP2022231,066L

259、arge language models on many tasks,particularly answering questions about specialized subjects like science and the humanities,such as logical reasoning and mathematics.BLOOMPA2022117433Multilingual and open source,the Bloom model,which has emerged from the BigScience participatory project,aims to h

260、elp advance research work on large language modelsChatGPT-4NLP202310062,318.8An advanced version of OpenAIs ChatGPT series,designed for more nuanced and context-aware language generation.OPT-175BNLP202333324A state-of-the-art language model by Meta known for generating coherent and contextually rele

261、vant sentences over long passages.1432225331|Powering Intelligence May 2024REFERENCES1.Statista.(2023).Data Center Market in the United States-Statistics&Facts.Statista.Link2.PreScouter.(2023).Data Center Subject Matter Expert Interviews for EPRI Research.PreScouter Interviews,2023(1).3.U.S.Departme

262、nt of Energy.(2017).Small Data Centers,Big Energy Savings:An Introduction for Owners and Op-erators-Final Report.U.S.Department of Energy:Better Buildings,AR(17).4.CBRE.(2023).Global Data Center Trends 2023.CBRE Research:Intelligent Investment Report,2023(1).Link5.Koomey,J.,Brill,K.,Turner,P.,Stanle

263、y,J.,&Taylor,B.(2007).White Paper:A Simple Model for Determining True Total Cost of Ownership for Data Centers.Uptime Institute,Inc.White Papers,1.6.Hoosain,M.S.,et al.(2023).Tools Towards the Sustain-ability and Circularity of Data Centers.Circular Economy and Sustainability,2023(3),173-197.7.Strub

264、ell,E.,Ganesh,A.,&McCallum,A.(2019).Energy and Policy Considerations for Deep Learning in NLP.Pro-ceedings of the 57th Annual Meeting of the Association for Computational Linguistics,36453650.8.Cushman&Wakefield.(2023).Global Data Center Market Comparison.Cushman&Wakefields Data Center Advi-sory Gro

265、up,2023(1).Link9.U.S.Energy Information Administration.(2023).Energy Information Administration,Electric Power Annual 2023.Link10.Andrae,A.S.(2020).New perspectives on internet electricity use in 2030.Engineering and Applied Science Letters,3(2),19-31.11.Koot,M.,&Wijnhoven,F.(2021).Usage impact on d

266、ata center electricity needs:A system dynamic forecasting model.Applied Energy,291(116798).12.Rahmani,R.,Moser,I.,&Seyedmahmoudian,M.(2017).A Complete Model for Modular Simulation of Data Centre Power Load.Journal of IEEE Transactions on Automation Science and Engineering,Vol.14(8).13.Shehabi,A.,Smi

267、th,S.J.,Sartor,D.,Brown,R.E.,Herrlin,M.,Koomey,J.,.&Lintner,W.(2016).United States data center energy usage report.Lawrence Berkeley National Laboratory,Berkeley,CA.14.Mysore,M.,Woetzel,J.,&Gupta,S.(2022).Investing in the Rising Data Center Economy.McKinsey&Company.Link15.Koomey,J.(2008).Worldwide e

268、lectricity used in data centers.Environmental Research Letters,3(034008).16.PJM Resource Adequacy Planning Department.(2022).2023 Load Forecast Supplement.PJM Resource Ad-equacy Planning Department.Link17.International Energy Agency(IEA).(2022).World En-ergy Outlook:2022.IEA:Annual World Energy Outl

269、ook Reports,2022(1).Link18.Shehabi,A.,Smith,S.J.,Masanet,E.,&Koomey,J.(2018).Data center growth in the United States:de-coupling the demand for services from electricity use.Environmental Research Letters,13(124030).19.TeleGeography.(2023).The State of the Network:2023 Edition.TeleGeography Annual R

270、eports,2023(1).Link20.Malmodin,J.,et al.(2023).ICT sector electricity con-sumption and greenhouse gas emissions:2020 out-come.SSRN,2023(1).21.CISCO.(2022).2022 Global Hybrid Cloud Trends Report.CISCO:Annual Global Hybrid Cloud Reports,2022(1).Link22.Khanboubi,Y.E.,&Hanoune,M.(2019).Exploiting Blockc

271、hains to improve Data Upload and Storage in the Cloud.International Journal of Communication Networks and Information Security(IJCNIS),11(3),357-364.23.Daigle,B.(2021).Data Centers Around the World:A Quick Look.United States International Trade Commis-sion:Executive Briefings on Trade,2(1).24.Andrae

272、,A.S.,&Edler,T.(2015).On global electricity usage of communication technology:trends to 2030.Challenges,6(1),117-157.25.Amazon.(2022).Annual Corporate Sustainability Re-port.Amazon Sustainability Reports,2022(1).Link26.Avelar,V.,et al.(2012).PUE:A Comprehensive Exami-nation of the Metric.The Green G

273、rid,White Paper 49(1).27.Verdecchia,R.,Sallou,J.,&Cruz,L.(2023).A systemic 1432225332|Powering Intelligence May 2024review of Green AI.WIREs Data Mining Knowledge Discovery,13(1507).28.Tibaldi,M.,&Pilato,C.(2023).A Survey of FPGA Op-timization Methods for Data Center Energy Efficiency.EEE Transactio

274、ns on Sustainable Computing,8(3),343-362.29.Davis,J.,et al.(2022).Uptime Institute:Global Data Center Survey 2023.Uptime Institute:Planning&Strat-egy UI Intelligence Report,78(1).30.EIRGRID&SONI.(2022).Ireland Capacity Outlook:2022-2031.Electric Power Transmission Operator in Ireland&System Operator

275、 for Northern Ireland,AR(22)31.Miller,R.(2023).Virginia State Legislators Target Data Center Development with New Bills.Data Center Fron-tier:Special Reports.Link32.CBRE.(2023).Dallas-Fort Worth Records Unprecedent-ed Data Center leasing Activity in First Half of 2023.CBRE Group,Inc.Link33.Sodhi,R.(

276、2023).How Californias New Emissions Dis-closure Law Will Affect Data Centers.Security Bloggers Network.Link34.Patel,D.,&Ahmad,A.(2023).The Inference Cost of Search Disruption Large Language Model Cost Analy-sis.SemiAnalysis Reports,2(1).Link35.DCF Staff.(2022).Why Phoenix is an Increasingly Hot Data

277、 Center Market.Data Center Frontier:Special Re-ports.Link36.JetCool Technologies,Inc.(2023).Drive Faster Com-puter Sustainability with Microconvective Cooling:How Microconvective Cooling Technology Provides Future-Ready Flexibility Meeting Data Centers Where They Are Today.JetCool Technologies,Inc.W

278、hite Papers,1.Doesnt have an integrated link in the original white pa-per,but here is a link to where you can find the content online.Link37.Miller,R.(2023).Atlanta Prepares for Data Center Building Boom Amid Growing Interest from Hyperscale Users.Data Center Frontier:Special Reports.Link38.Desislav

279、ov,R.,Martinez-Plumed,F.,&Hernandez-Oral-lo,J.(2023).Trends in AI Interface Energy Consump-tion:Beyond the Performance-vs-Parameter Laws of Deep Learning.Sustainable Computing:Informatics and Systems,38(100857).39.Stanford University:Human-Centered Artificial Intelli-gence.(2023).Artificial Intellig

280、ence Index Report 2023.Stanford University:HAI.Link40.Costenaro,D.,&Duer,A.(2012).The Megawatts be-hind Your Megabytes:Going from Data-Center to Desk-top.American Council for an Energy-Efficient Economy.41.Imperva.(2023).2023 Imperva Bad Bot Report.Im-perva Bot Reports,2023(1).Link42.National Renewa

281、ble Energy Laboratory.(2022).2022 Standard Scenarios Report:A U.S.Electricity Sector Outlook.National Renewable Energy Laboratory.Link43.Mittal,S.(2019).A Survey of Techniques for Improv-ing Energy Efficiency in Embedded Machine Learning Systems.arXiv Preprint,arXiv:1904.10462.44.Patterson,D.,et al.

282、(2022).The Carbon Footprint of Machine Learning Training Will Plateau,Then Shrink.Google&University of California-Berkeley.Link45.SimilarWeb.(2023).ChatGPT and Competitors:Weekly Visits Desktop&Mobile Web,U.S.SimilarWeb Data Comparison,2(1).46.OECD.(2022).Measuring the Environmental Impacts of Artif

283、icial Intelligence Compute and Applications:The AI Footprint.OECD Digital Economy Papers:OECD Publish-ing,2022(341).47.Vries,A.D.(2023).The Growing Energy Footprint of Artificial Intelligence.Joule,7(1),2191-2194.48.Reid,E,et al.(2023).Supercharging Search with Gen-erative AI.Google Blog Products.Li

284、nk49.U.S.Environmental Protection Agency.(2022).16 More Ways to Cut Energy Waste in the Data Center.ENERGY STAR.Link50.National Renewable Energy Laboratory.(2017).The Uniform Methods Project:Methods for Determining Energy Efficiency Savings for Specific Measures.Na-tional Renewable Energy Laboratory

285、.Link51.Xu,W.,et al.(2021).Accelerating Federated Learning for IoT in Big Data Analytics with Pruning,Quantization and Selective Updating.IEEE,9(1),38457-38766.52.Jouppi,N.,&Patterson,D.(2023).Googles Cloud TPU v4 provides exaFLOPS-scale ML with industry-leading 1432225333|Powering Intelligence May

286、2024efficiency.Google Cloud Products.Link53.Nguyen,T.,et al.(2020).The Performance and Energy Efficiency Potential of FPGAs in Scientific Computing.Lawrence Berkeley National Laboratory(LBNL).54.Qasaimeh,M.,et al.(2019).Comparing Energy Efficien-cy of CPU,GPU and FPGA Implementations for Vision Kern

287、als.IEEE,978(1),7281-7289.55.Heydari,A.,et al.(2022).Power Usage Effectiveness Analysis of a High-Density Air-Liquid Hybrid Cooled Data Center.Proceedings of the ASME 2022 Interna-tional Technical Conference and Exhibition on Packag-ing and Integration of Electronic and Photonic Micro-systems,25(27)

288、.56.JetCool Technologies,Inc.(2023).Drive Faster Com-puter Sustainability with Microconvective Cooling:How Microconvective Cooling Technology Provides Future-Ready Flexibility Meeting Data Centers Where They Are Today.JetCool Technologies,Inc.White Papers,1.57.Meta.(2022).OCP Summit 2022:Grand Teton

289、.Face-book Engineering.Link58.Page Southerland Page,Inc.(2022).White Paper:Two-Phase Liquid Immersion Cooling Case Study.Page Southerland Page White Papers,1.59.SkyCool Systems.(2019).Harnessing the Cold of the Sky and Space to Enable Electricity-free Cooling.Sky-Cool Systems:Radiative Cooling Techn

290、ology Case Study.Link60.Submer&Wyoming Hyperscale.(2022).Immersion Cooling for Hyperscalers:Powering Farming of the Future.Submer White Papers&Wyoming Hyperscale White Box,1.Link61.Texas Advanced Computing Center&Green Revolution Cooling.(2023).Advanced Cooling Advances Science Case Study.TACC Case

291、Studies,1.62.Ham,S.W.,Kim,M.H.,Choi,B.N.,&Jeong,J.W.(2015).Energy saving potential of various air-side economizers in a modular data center.Applied Energy,135(15),258-275.63.Barroso,L.A.,&Hlzle,U.(2007).The case for energy-proportional computing.Computer,40(12).64.U.S.Environmental Protection Agency

292、.(2022).Annual Emissions Report by Industry Sector.U.S.Environmen-tal Protection Agency.Link65.Apple.2022 Environmental Progress Report.Apple Sustainability Reports,2022(1).Link66.Google.(2022).Environmental Report.Google Sustain-ability Reports,2022(1).Link67.Meta.2022 Sustainability Report:For a B

293、etter Reality.Meta Sustainability Reports,2022(1).Link68.Microsoft.2022 Environmental Sustainability Report.Microsoft Sustainability Reports,2022(1).Link69.Larson,A.(2023).NuScale Gets a Win with SMRs for Data Centers in Ohio and Pennsylvania.PowerMag Publishing.Link70.Chhabra,S.,&Singh,A.(2023).Dyn

294、amic Resource Allocation Method for Load Balance Scheduling over Data Center Networks.Journal of Web Engineering,2211(02352).71.Ghatikar,G.,et al.(2012).Demand Response Opportu-nities and Enabling Technologies for Data Centers:Find-ings from Field Studies.Lawrence Berkeley National Laboratory(LBNL).

295、72.Mehra,V.,&Hasegawa,R.(2023).Supporting power grids with demand response at Google Data Centers.Google Cloud Products.Link73.Abhyankar,N.,et al.(2021).2030 Report:Powering Americas Clean Economy.University of California-Berkeley:Goldman School of Public Policy,2021(1).74.Horner,N.,Shehabi,A.,&Azev

296、edo,I.(2016).Known unknowns:Indirect energy effects of information and communication technology.Environmental Research Letters,11(10).75.Raizada,A.,&Singh,K.(2020).Worldwide energy consumption of hyperscale data centers:a Survey.In-ternational Research Journal on Advanced Science Hub,02(11).76.U.S.E

297、nvironmental Protection Agency.(2018).Quanti-fying the Multiple Benefits of Energy Efficiency and Re-newable Energy.U.S.Environmental Protection Agency.Link77.Howard,A.(2022).Data Center Building Report 2022.OMDIA:Annual Reports,2(1).Link78.Data Center Frontier.(2023).The Power Problem:1432225334|Po

298、wering Intelligence May 2024Transmission Issues Slow Data Center Growth.Data Center Frontier:Special Reports.Link79.Masanet,E.,et al.(2020).Recalibrating global data cen-ter energy-use estimates.Science(AAAS),367(6481),984-986.80.Mytton,D.,&Ashtine,M.(2022).Sources of data cen-ter energy estimates:A

299、 comprehensive review.Joule,6(1),20322056.81.Movva,R.,Lei,J.,Longpre,S.,Gupta,A.,&DuBois,C.(2022).Combining Compressions for Multiplicative Size Scaling on Natural Language Tasks.Proceedings of the 29th International Conference on Computational Linguistics.82.Sevilla,J.,et al.(2022).Compute Trends A

300、cross Three Eras of Machine Learning.2022 International Joint Con-ference on Neural Networks(IJCNN),1(8).83.Bastian,M.(2023).GPT-4 Has More Than a Trillion Parameters Report.The Decoder Reports,1.Link84.BigScience.(2022).Introducing the Worlds Largest Open-source Multilingual Language Model:BLOOM.Bi

301、gScience Analysis,1.Link85.DeepChecks.(2023).LLM Models Comparison:GPT-4,Bard,LLaMA,Flan-UL2,BLOOM.DeepChecks Analysis,1.Link86.Hoffman,J.,et al.(2023).Training Compute-Opti-mal Large Language Models.arXiv Preprint,arX-iv:2203.15556.87.Rae,J.,Irving,G.,&Weidinger,L.(2021).Language Modelling at Scale

302、:Gopher,Ethical Considerations,and Retrieval.Google DeepMind Research,1.88.Zhang,S.,et al.(2023).OPT:Open Pre-trained Transformer Language Models.arXiv Preprint,arX-iv:2205.0106814322253About EPRIFounded in 1972,EPRI is the worlds preeminent independent,non-profit energy research and development org

303、anization,with offices around the world.EPRIs trusted experts collaborate with more than 450 companies in 45 countries,driving innovation to ensure the public has clean,safe,reliable,affordable,and equitable access to electricity across the globe.Together,we are shaping the future of energy.EPRI3420

304、 Hillview Avenue,Palo Alto,California 94304-1338 USA 650.855.2121 2024 Electric Power Research Institute(EPRI),Inc.All rights reserved.Electric Power Research Institute,EPRI,and TOGETHERSHAPING THE FUTURE OF ENERGY are registered marks of the Electric Power Research Institute,Inc.in the U.S.and worl

305、dwide.3002028905 May 2024EPRI Customer Assistance Center 800.313.3774 For more information,contact:PRINCIPAL INVESTIGATORSJORDAN ALJBOUR,Strategic Insight Engineer TOM WILSON,Executive Technical Leader PRINCIPAL INVESTIGATOR AND EPRI CONTACTPOORVI PATEL,Manager Strategic Insight 704.232.4551,14322253