GPU计算

AmesLaboratoryTamingComplexity:ThePromiseofHighPerformanceComputingBESScience2018:AFutureRetrospectiveSymposiuminthestyleofIranThomasBruceHarmonComplexityAlmostZen:Ifyoutameacomplexproblem,itisnolongeracomplexproblem.“Ihaveyettoseeanyproblem,howevercomplicated,whichwhenyoulookedatitintherightway,didnotbecomestillmorecomplicated.”PoulAndersonPrinciplefor(Successful)Predictions“Parisinthetwentiethcentury-and,morespecifically,in1960-teemswithprodigiesthatwerehardtoimagineahundredyearsbefore;yetJulesVerneimaginedthembecausethescienceandtechnologyofhisdaysuggestedtheirpossibility.Imaginationisthecapacitytorearrangeavailabledataortoextrapolatefromthem,andVernewasamasterfulextrapolator.”FromEugenWeber’sintroductionto“ParisintheTwentiethCentury”writtenin1863byJulesVerne,butnotfounduntil1989.Moore’sLaw•Itcannotcontinue.Therewillnotbepicotechnology.•Thequestionis“howfarcanwepush?”•Manyrationalargumentsforpessimism.•Optimismrequiresbeliefthatdifficultproblemscanbeovercome…•2018isofftheroadmap,butcleverpeoplecandoamazingthings(notmagic).SomePastPredictions:1943Ithinkthereisaworldmarketformaybefivecomputers.--ThomasWatson,chairmanofIBM1949Computersinthefuturemayweighnomorethan1.5tons.”--PopularMechanics,forecastingtherelentlessmarchofscience1977Thereisnoreasonanyonewouldwantacomputerintheirhome.--KenOlson,president,chairmanandfounderofDigitalEquipmentCorp.1981“640Koughttobeenoughforanybody”--BillGatesDramaticadvancesintheoryandalgorithms•InComputationalMaterialsSciencetheamplificationfromintellect(newtheory,algorithms,andinsights)isatleastequaltothespeed-upfromhardwareRelativeperformanceincreaseofIsingmodelsimulations(■)comparedthenormalizedspeedofthecomputers(●)thesimulationswereexecutedon.Thedashedlineisaschematicoftheincreaseinpeakperformanceofthefastestsupercomputerssince1972.197019751980198519901995200011010010001000010000010000001E71E81E91E10relativeperformancecomputerspeedMoore’slawDavidLandau:UGAManyCMSapplicationcodesscaleonparallelcomputersCompaq,PSC4.58Tflops!xSimilardataforothercodesParatecFLAPWGordonBellPrizes1992–KKR-CPA:PricePerformance1995-TBMD:Fastestrealapplication1999–LSMS:Fastestrealapplication104improvementsince1988GBPrize1988-1Gflop/s2001-11Tflop/sDramaticAdvancesinTheoryandAlgorithms•Allowtreatmentofcomplexsystemsandphenomena•Acoupleofexamples/projections:–Lightweightmaterials•MicrostructureinAluminumcasting–Magneticstorageandrecording•Nano-magnetismMicrostructure–PropertyRelationshipsEngineBlock1meterMicrostructure-Grains1–10mmProperties•Highcyclefatigue•DuctilityMicrostructure-Phases100–500micronsProperties•Yieldstrength•Ultimatetensilestrength•Highcyclefatigue•Lowcyclefatigue•ThermalGrowth•DuctilityMicrostructure-Phases3-100nanometersProperties•Yieldstrength•Ultimatetensilestrength•Lowcyclefatigue•DuctilityChrisWolvertonFordMotorCompanyAtoms10-100AngstromsProperties•ThermalGrowthActiveStorageDensityTrendsDensity,Mbits/in2Year199020002010100,00010,0001,00010010MagneticStorageDensitySemiconductorStorageDensityNSF-MRSEC05/22/03ISUKECKCMRAMagneticNanostructuresIBMdiskdrivewithGBcapacityGMRReadHeadTechnologydemandstreatmentofcomplexphenomenaGMRreadheadtakesadvantageofadvancesinbasicscienceOscillatoryexchangecouplingGiantMagneto-ResistanceExchangeBiasCuNi-FeMagneticdomainsMediamotionReader:Spin-valveCoWriterReaderBottomShieldMediaImportanceofsurfacesandinterfacesdemandsfirstprinciplesapproachesSynthesis&CharacterizationTheory&SimulationNanowiresandparticlesMagneticstructureofmatrixandsurfacenanostructuresDynamics&switchingSpintransportMagneticnanostructuresEffectsofDimensionalConfinementonSpinStructure,DynamicsandTransportSpintronicsFirstprinciplessimulationsizeRealsystemsizeNanoscience–ComplexitydemandsModeling5nmFe~12,000atoms~2,000surfaceatomsEffectsofDimensionalConfinementofSpinStructure,DynamicsandTransportNanowiresandparticlesConfinementmodifiesSpinstructure,Dynamics,TransportDecoherenceandQCNewconceptsActivecontrolofspin2018:NanomagnetismQuantumCorral:Eigler(IBM)NewQuantumPhenomenaMagneticMoleculesasBuildingBlocksFe30AnEmergingTransformationintheStructureofComputationalMaterialsScience•Historicalstructureoftheoryandcomputationincondensedmatterphysics–Manycomplexproblems•Nooverarching“problem”–Rapidlychangingfocus–Rapidlydevelopingtechniques–Codes/Methodsdevelopedbysmallgroups-students•“CottageIndustry”–BillMcCurdy–Sophisticatedmodelsandcodesexistateverylengthscale.–Thesecodescomeoutofdifferentcottagesandarehardtointegrate.–Developedforincompatiblearchitectures.–Fastpaceofdevelopmentdoesnotallowforrewritingcodesfromscratch.QuantumSpinModelsMicromagneticsMicromagneticsLargeFORTRANcodes(20,000lines)developedoveralongtimeperiod(20years)MonolithicCorFORTRANcodesusuallydesignedtosolveaparticularproblemCodesdesignedtorunonPCswithemphasisongraphicaluserinterfaceUltrascaleComputationdemandsanewapproachNewApproach:NanoScienceJointBESAC-ASCACNanoScienceworkshop–May2002FullydevelopingthepotentialoftheDOE-OSNano-centerswillrequireanewlevelofintegrationofexperimen