MLPerf基准套件技术概述.pptxVIP

DemystifyingMLPerfBenchmarkSuiteMLPerf基准套件技术概述

AgendaTargetAudienceAboutusMLPerfintroductionHowtostart

TargetAudienceAnyonewhohasinterestonDeepLearningcodingAnyonewhowanttohaveareferenceorbaselineforsystempurchaseGPUCPU/MemoryFPGAandothersCloudSolutionprovidersEX.Dell,publishingbenchmarksonserver/storageAnsweringRFPsAnyonewholiketocontributeinMLPerfcommunity

World-classinfrastructureintheInnovationLabZenithTOP500-classsystembasedonIntelScalableSystemsFramework(OPA,KNL,Xeon,OpenHPC)424nodesdualIntelXeonGoldprocessors,Omni-Pathfabric.?+160IntelXeonPhi(KNL)servers.Over1PFcombinedperformance!#265onTop500June2018,1.86PFtheoreticalpeakLustre,IsilonH600,IsilonF800andNSSstorageLiquidcooledandaircooledRattlerResearch/developmentsystemwithMellanox,NVIDIAandBrightComputing88nodeswithEDRInfiniBandandIntelXeonGoldprocessors32xPowerEdgeC4140nodeswith4xNVIDIAGPUsOthersystems32nodeAMDcluster,storagesolutions,etc.13Kft.2lab,1,300+servers,~10PBstoragededicatedtoHPCincollaborationwiththecommunity

HPCandDLEngineering-whatwedoDesignandbuildsystemsforHPCandDeepLearningworkloads.Systemsincludecompute,storage,network,software,services,support.Integrationwithfactory,software,services.Powerandperformanceanalysis,tuning,bestpractices,trade-offs.Focusonapplicationperformance.Verticalsolutions.Researchandproofofconceptstudies.Publishwhitepapers,blogs,conferencepapers()Accesstothesystemsinthelab

MLPerfintroductionWhatisMLPerfAopensourcedMLbenchmarksuiteformeasuringperformanceofMLframeworks,MLhardwareaccelerators,andMLcloudplatforms.CoverdifferentDLdomainsPropermetrics(trainingtime,accuracy,latency)RealdatasetsWhatisthegoalofMLPerfFairandusefulbenchmarksformeasuringtrainingandinferenceperformanceofMLhardware,software,andservices.WhatisthetypesofML