- 1、本文档共15页,可阅读全部内容。
- 2、原创力文档(book118)网站文档一经付费(服务费),不意味着购买了该文档的版权,仅供个人/单位学习、研究之用,不得用于商业用途,未经授权,严禁复制、发行、汇编、翻译或者网络传播等,侵权必究。
- 3、本站所有内容均由合作方或网友上传,本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺!文档内容仅供研究参考,付费前请自行鉴别。如您付费,意味着您自己接受本站规则且自行承担风险,本站不退款、不进行额外附加服务;查看《如何避免下载的几个坑》。如果您已付费下载过本站文档,您可以点击 这里二次下载。
- 4、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等,请点击“版权申诉”(推荐),也可以打举报电话:400-050-0827(电话支持时间:9:00-18:30)。
JournalofMachineLearningResearch24(2023)1-15Submitted1/23;Revised5/23;Published6/23
EstimatingtheCarbonFootprintofBLOOM,
a176BParameterLanguageModel
AlexandraSashaLuccioniSASHA.LUCCIONI@HUGGINGFACE.CO
HuggingFace
Montréal,Canada
SylvainViguierSYLVAINV@GRAPHCORE.AI
Graphcore
London,UK
Anne-LaureLigozatANNE-LAURE.LIGOZAT@LISN.UPSACLAY.FR
LISNENSIIE
Paris,France
Editor:ShakirMohamed
Abstract
Progressinmachinelearning(ML)comeswithacosttotheenvironment,giventhattrainingML
modelsrequirescomputationalresources,energyandmaterials.Inthepresentarticle,weaimto
quantifythecarbonfootprintofBLOOM,a176-billionparameterlanguagemodel,acrossitslife
cycle.WeestimatethatBLOOM’sfinaltrainingemittedapproximately24.7tonnesofCOeqif
2
weconsideronlythedynamicpowerconsumption,and50.5tonnesifweaccountforallprocesses
rangingfromequipmentmanufacturingtoenergy-basedoperationalconsumption.Wealsocarryout
anempiricalstudytomeasuretheenergyrequirementsandcarbonemissionsofitsdeploymentfor
inferenceviaanAPIendpointreceivinguserqueriesinreal-time.Weconcludewithadiscussion
regardingthedifficultyofpreciselyestimatingthecarbonfootprintofMLmodelsandfutureresearch
directionsthatcancontributetowardsimprovingcarbonemissionsreporting.
Keywords:carbonfootprint,languagemodeling,lifecycleassessment,machinelearning.
1.Introduction
Climatechange
文档评论(0)