gpu architecture overview

《gpu architecture overview》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、GGPPUUAArrchichittectecturureeOOvverervviiewewJoJohhnnOOwwenenssUUCCDavDaviissTheTheRiRightght--HanHanddTurTurnn[H&PFigure1.1]WWhy?hy?[A[ArrchichittectecturureeReasoReasonns]s]•ILPincreasinglydifficulttoextractfrominstructionstream•Controlhardwaredominatesµprocessors–

2、Complex,difficulttobuildandverify–Takessubstantialfractionofdie–Scalespoorly•Payformaxthroughput,sustainaveragethroughput•Quadraticdependencychecking–Controlhardwaredoesn’tdoanymath!•IntelCoreDuo:48GFLOPS,~10GB/s•NVIDIAG80:330GFLOPS,80+GB/s3AAMMDD““DDeereerhohounundd”

3、”((KK88L)L)chip-architect.com4WWhy?hy?[Techn[TechnoolologygyReasoReasonns]s]•Industrymovingfrom“instructionspersecond”to“instructionsperwatt”–“Powerwall”nowall-important–Traditionalµproctechniquesarenotpower-efficient•Wecancontinuetoputmoretransistorsonachip…–…butweca

4、n’tscaletheirvoltagelikeweusedto…–…andwecan’tclockthemasfast…5GGooPPararallelallel•Timeofarchitecturalinnovation–GPUsletusexploreusinghundredsofprocessorsnow,not10yearsfromnow•MajorCPUvendorssupportingmulticore•Interestingeneral-purposeprogrammabilityonGPUs•Universiti

5、esmustteachthinkinginparallelWWhathat’’ssDiDifferfferenenttaboaboututttheheGGPPU?U?•Thefutureofthedesktopisparallel–Wejustdon’tknowwhatkindofparallel•GPUsandmulticorearedifferent–Multicore:Coarse,heavyweightthreads,betterperformanceperthread–GPUs:Fine,lightweightthrea

6、ds,single-threadperformanceispoor•AcasefortheGPU–Interactionwiththeworldisvisual–GPUshaveawell-establishedprogrammingmodel–MarketforGPUsis500M+total/year7TheTheRenRenderderiinnggPPiipelipelinneeApplicationCompute3DgeometryMakecallstographicsAPITransformgeometryfrom3Dt

7、oGeometry2D(inparallel)Generatefragmentsfrom2DRasterizationgeometry(inparallel)CompositeCombinefragmentsintoimageGPUTheThePPrroogrgramammmableablePPiipelipelinneeApplicationCompute3DgeometryMakecallstographicsAPITransformgeometryfrom3DtoGeometry2D[vertexprograms]Gener

8、atefragmentsfrom2DRasterizationgeometry[fragmentprograms]CompositeCombinefragmentsintoimageGPUDiDirrectectXX1100PPiipelipeli