邓仰东：基于GPU的高性能嵌入式计算CUDA技术沙龙

《邓仰东：基于GPU的高性能嵌入式计算CUDA技术沙龙》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、HighPerformanceEmbeddedComputingwithMassivelyParallelProcessorsYangdongSteveDeng邓仰东dengyd@tsinghua.edu.cnTsinghuaUniversityOutlineMotivationandbackgroundMorphingGPUintoanetworkprocessorHighperformanceradarDSPprocessorConclusionHighPerformanceEmbeddedComputingFutureITi

2、nfrastructuredemandsevenhighercomputingpowerCoreInternetrouterthroughput:upto90Tbps4Gwirelessbasestation:1Gbit/sdataratepercustomerandupto200subscribersinserviceareaCMUdriverlesscar:270GFLOPs(GigaFLoatingpointOperationsPersecond)…~$1MFastIncreasingICCostsFabricationCo

3、stMoore’sSecondLaw:ThecostofdoublingcircuitdensityincreasesinlinewithMoore'sFirstLaw.DesignCostNow$20-50MperproductWillreach$75-120Mat32nmnodeThe4-yeardevelopmentofCellprocessorbySony,IBM,andToshibacostsover$400M.ImplicationsoftheProhibitiveCostASICswouldbeunaffordabl

4、eformanyapplications!ScottMacGregor,CEOofBroadcom:“Broadcomisnotintendingamoveto45nminthenextyearorsoasitwillbetooexpensive.”DavidTurek,VPofIBM:“IBMwillbepullingoutofCelldevelopment,withPowerXCell8itobethecompany’slastentranceinthetechnology.”MulticoreMachinesAreReall

5、yPowerful!ManufacturerProcessorTypeModelModelNumber#CoresGFLOPsFP64GFLOPsFP32AMDGPGPUFireStream9270160/8002401200AMDGPURadeonHD5870320/16005442720AMDGPURadeonHD5970640/32009284640AMDCPUMagny-Cours12362.11362.11FujitsuCPUSPARC64VII4128128IntelCPUCore2ExtremeQX9775451.2

6、51.2nVidiaGPUFermi4805127801560nVidiaGPGPUTeslaC106024077.76933.12nVidiaGPGPUTeslaC2050448515.21288TileraCPUTilePro64166166AMD12-CoreCPUTileraTileGx100CPUNVidiaFermiGPUGPU:GraphicsProcessingUnitGPGPU:GeneralPurposeGPUImplicationsAnincreasingnumberofapplicationswouldbe

7、implementedwithmulti-coredevicesHuawei:multi-corebasestationsIntel:clusterbasedInternetroutersIBM:signalprocessingandradarapplicationsonCellprocessor…AlsomeetsthestrongdemandsforcustomizabilityandextendibilityOutlineMotivationandbackgroundMorphingGPUintoanetworkproces

8、sorHighperformanceradarDSPprocessorConclusionBackgroundandmotivationGPUbasedroutingprocessingRoutingtablelookupPacketclassif