AMD先进封装技术——过去、现在和未来.pdf

1、AMDAMD Advanced Advanced Packaging Packaging Past,Past,Present,and FuturePresent,and FutureAMD HIT(Heterogeneous Integrated Technologies)AMD Advanced Packaging Past,Present,and FutureDevin WuPrincipal Member of Technical Staff Heterogeneous Integration TechnologiesAMDCHIPLETS AND ADVANCED PACKAGING/

2、PHOTONICSOutline4321IntroductionAdvanced Packaging AMD LeadershipAMD Chiplet TechnologyFuture DirectionsAI Driving Compute&Memory GrowthCompute Requirements Petaflops/dayCompute Requirements Petaflops/dayDoubling every 3.4 monthsMoores Law doubling every 24 months201220122013201320142014201520152016

3、201620172017201820181e+41e+21e+01e-21e-31e-41e-51e+31e+11e-1Source:Source:https:/ computeAI computeAlexnet to alphago zero:A 300,000X increase in compute1 exaFLOP1 petaFLOP1 gigaFLOP1990199019951995200020002005200520102010201520152025202520202020FrontierFrontierLeading supercomputer performanceLeadi

4、ng supercomputer performanceSource:Source:https:/www.top500.orghttps:/www.top500.orgParameters(log scale)Image,Speech ModelsBERTBERT-large large 330 million330 millionResNet50 ResNet50 26 million26 million2x2x/yearyearAmoebaNetB AmoebaNetB 557 million557 millionSwitch 1.6 Switch 1.6 trilliontrillion

5、20 x20 x/yearyear2014201420152015201620162017201720182018201920192021202120202020AI memoryAI memorySource:Source:TechInsightsTechInsights(2022)(2022)AMD:Broadest AI Product PortfolioCloud and Cloud and EnterpriseEnterpriseGeneral AI InferenceEdge/Edge/EmbeddedEmbeddedEmbedded InferenceCloud and Clou

6、d and SupercomputerSupercomputerHPC and AI TrainingDense InferenceEndpointEndpointAI Inference for PCsGaming AIGaming AIGaming AIAI Enabled:Cloud to edge to endpointAI Enabled:Cloud to edge to endpointDiverse Requirements for AICloud/EnterpriseCloud/Enterprise1000+of TFLOPS computeTCO,demanding comp