可持续生成式 AI 构建：从理念到落地的完整指南.pptxVIP

Developing

sustainableGenAI

REEARHNTTUTE

#GetTheFutuΓeYouWant

Tableof

contents

ExecutivesummaΓy

09

GenAIhas

asignificant

environmentalimpact

ThesustainabilityofGenAIremainsalowpriority

Whoshouldread

thisreportandwhy?

04

10

18

DevelopingsustainableGenAI

2

Aroadmapto

responsibleGenAIforsustainablebusinessvalue

Conclusion

65

ReseaΓchmethodology

60

32

DevelopingsustainableGenAI

3

GenerativeAIhasasignificantenvironmentalimpact:

GenerativeAI(GenAI)reliesonprocessinghugevolumesofdata,whichnecessitatesenormouscomputationalpower,

renderingitanenergy-intensivetechnology.Theproductionofgraphicsprocessingunits(GPUs),integraltothe

functioningofGenAI,requiresrareearthmetals,theminingofwhichcontributestogreenhousegas(GHG)emissions.

Additionally,thefrequenthardwareupgradesthatGenAIrequiresputagreatdealofstressonnaturalresources,aswellasfurtherpollutingtheenvironment.Recentestimatessuggestthat,GenAIcouldcreatebetween1.2to5.0millionmetrictonsofe-wasteby2030,whichisaround1,000timesmoree-wastethanwasproducedin2023.1

EstimatessuggestthattrainingaGPT-3model(which

includes175billionparameters)consumesanamountof

electricityequivalenttotheannualconsumptionof130

UShomes.2Movingtothenextmodelsizeup,GPT-4(with

1.76trillionparameters),powerconsumptionoftrainingisestimatedtobeequivalenttoyearlypowerconsumptionof5,000UShomes.3Aftertraining,whichisaone-timeeventinthemodel’slifecycle,theinferencingphase,wheremodels

performreal-timefunctions,requiresanequalorgreater

amountofenergy.Datacentersalsoconsumeahugeamountofenergyandwaterforcoolingpurposes.Runningan

inferenceof20–50queriesonanLLMusesabout500mlofwatereachtime.

GenAIhascontributedtoincreasedGHGemissions:

Inourresearch,wesurveyedexecutivesfrom2,000organ