DC25_SLIDES_Chiphead_PanelEnergyEfficientInterfacesFor_Hutchins.pdf

1、Information Classification:GeneralWelcome to ConferenceJanuary 2830,2025Santa Clara Convention Center1ExpoJanuary 2930,2025 Information Classification:GeneralEnergy Efficient Interfaces for the Next Generation of AI ComputePanel Discussion:Jeff Hutchins,Moderator(OIF and Ranovus)Nathan Tracy(OIF and

2、 TE Connectivity)Mike Klempa(OIF and Alphawave Semi)Sam Kocsis(OIF and Amphenol)2 2Information Classification:GeneralSPEAKERSJeff HutchinsCTO Office,Director,ROIF Physical&Link Layer Working Group-EEI Vice Chair&OIF TreasurerMike KlempaProduct Marketing,Alphawave SemiMichael.K|OIF Board Member 3Imag

3、eImage Nathan TracyTechnologist,System Architecture TE COIF President of the BoardSam KocsisDirectory of Standards,Amphenolsamuel.kocsisamphenol-OIF Technical Committee Vice Chair3DesignCon 2025 Energy Efficient Interfaces for the Next Generation of AI Compute 4AGENDADesignCon 2025 Energy Efficient

4、Interfaces for the Next Generation of AI Compute 51Overview of OIFs work on Energy Efficient Interfaces(EEI)Jeff Hutchins,RanovusOIF PLL WG EEI Vice Chair2Energy Efficient Ethernet InterfacesMike Klempa,Alphawave SemiOIF Board MemberOIF PLL Interoperability WG Chair3Energy Efficient Local Compute In

5、terfacesSam Kocsis,AmphenolOIF PLL Vice Chair4Exploring the next data rate:448GNathan Tracy,TE ConnectivityOIF President5Q/A Panel DiscussionThe ChallengeDesignCon 2025 Energy Efficient Interfaces for the Next Generation of AI Compute AI training utilizes large quantities of matrix multiplicationGPU

6、s are designed to accelerate“multiply and add”operations used in AI matrix computationEach row in matrix A is paired with every column in matrix B Lots of computation with lots of parameters!Large AI models can partition the computation into smaller chunksTile computations can be handed off to clust