GPU Acceleration

《GPU Acceleration》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、GPUAccelerationGPUandmixedCPU/GPUprocessingisintegratedinFEKOforrun-timecriticalphasesofthesolversutilisedbyFEKO.Themostrun-timecriticalsolutionphaseofmoderatetolargesizeMethodofMoments(MoM)problemsisthesolutionofthesystemoflinearequations.EventhoughFEKOemployshighlyoptimisedlibrariesforthispha

2、seforvariousCPUs,theGPUcapabilitiescanacceleratethisperformancebymorethananorderofmagnitude.FEKO'simplementationisuniqueasitalsoallowsthesolutionsofproblemswhichexceedthememoryofthegraphicscard(blockingwillbedonethen).Bothsingleanddoubleprecisioncomputationsaresupported.FEKOalsohastheabilitytou

3、semultipleGPUsinasinglesolutionrun,acceleratingsolutionspeedinrelationtothenumberofGPUsthatareapplied.Asanexample,considertheradarscatteringfromametallicobjectatasinglefrequency.ThefollowingfiguresgivetheperformanceinGFLOPS(billionsoffloatingpointoperationspersecond)ofthematrixsolutionphase(i.e

4、.excludingtimesformatrixsetup,near-andfar-fieldcalculationsetc.)forsolvingthisclassofproblemwiththeMoM.Measuredperformanceisgivenforproblemsizesrangingfrombetween2000and20000unknowns,withahigherperformancevalueindicatingthataproblemofacertainsizewillbesolvedmorequickly.1Figure1:GFLOPSperformanc

5、e(DoublePrecision)inFEKOfortheMoMsolutionphaseofsolvingthesystemoflinearequationsusingdifferentIntelCPUsandNVidiaGraphicsCards(GPUs).2Figure2:GFLOPSperformance(SinglePrecision)inFEKOfortheMoMsolutionphaseofsolvingthesystemoflinearequationsusingdifferentIntelCPUsandNVidiaGraphicsCards(GPUs).Thes

6、uperiorityofGPUprocessingforthematrixsolutionphaseisobvious.ResultsalsoshowthatthesingleprecisionsolutionisabouttwotimesfasterthandoubleprecisionandthattheadditionofextraGPUsimprovesperformancesignificantly.3