通过更高效地利用资源来推进人工智能工厂的发展.pdf

1、Advancing the AI Factory:Doing More with MoreCiscoEnterpriseAdvancing the AI Factory:Doing More with MoreAI CLUSTERChih-Tsung HuangSenior DirectorWei-Jen HuangDistinguished EngineerCiscoCiscoEnterprise100 Million Monthly Active UsersUberTelegramSpotifyPinterestInstagramTikTokChatGPT2 months9 months3

2、0 months41 months55 months61 months70 months35x TelegramSpotifyPinterestInstagramTikTok9 months30 months41 months55 months61 monthsWhere Are We Today?E X P E C T A T I O N ST I M EInnovation TriggerPeak of Inflated ExpectationsTrough of DisillusionmentSlope of EnlightenmentPlateau of ProductivityBui

3、ld the ModelCustom foundation models 100,1,000,10,000 GPU ClustersInfiniBand or EthernetWEB SCALERTraining/InferencingLARGE ENTERPRISETraining/InferencingBuild the ModelOptimize the ModelFine tuning pre-trained models 4-8 GPU Nodes EthernetINFERENCING ModerateTRAININGModerateOptimize the ModelUse th

4、e ModelPre-trained models with RAG2 GPU NodesEthernetINFERENCINGPre-trained ModelsINFERENCINGas-a-ServiceUse the ModelGenerative AI SpectrumInfrastructure RequirementsMost GenAI projects are hereDoing More with PowerMore We Use,More We Waste80Watts20Watts100 Watts199520302000200520102015202020253kW5

5、kW10kW15kW20kW30kW100kW200kW0.6kW1kW2kW3kW4kW6kW20kW40kW6010020030040060020004000Power Supply Efficiency ConsiderationPinPout1000 W800 W+DC-AC-DCInverter RectifierPower Factor Correction830 W=17Adopt the Highest Efficient Power Supplyhttps:/ 1 5 V92%1 1 5 V90%1 1 5 V88%1 1 5 V85%1 1 5 V96.5%80%1 1 5

6、 VTune Power Supply for Your Use Case80 PLUS Standard80 PLUS Bronze80 PLUS Silver80 PLUS Gold80 PLUS Platinum80 PLUS Titanium80 PLUS Ruby25%ENTERPRISE50%20%100%80 PLUS75%WEBSCALERDoing More with ThermalMore We Use,More We Waste ()3100%0%Fan PowerFan Speed1812Standard Proportional