Application-Specific Microarchitectural Customizat

CustomizableEmbeddedSystemArchitecturesPeterPetrovandAlexOrailogluUniversityofCalifornia,SanDiegoEmbeddedProcessorsMarketEmbeddedprocessorsoccupymorethan90%oftheentireprocessormarketAlargenumberofelectronicproductsrequirehigh-end32/64-bitsembeddedprocessorsEmbeddedProcessorsMarketEmbeddedprocessorsoccupymorethan90%oftheentireprocessormarketAlargenumberofelectronicproductsrequirehigh-end32/64-bitsembeddedprocessorsCellphones•DSPcomputation•Speechcodecs•WirelessprotocolsAutomotive•Enginecontrol•Automatictransmission•ABS,GPSMultimedia•DVD/MP3/CD•Videogames•DigitalcamerasApplicationRequirements•Designcost•Time-to-market•FlexibilityGeneralembeddedprocessorarchitecturesintroducedtosatisfytheseconstraints!ApplicationRequirements•Designcost•Time-to-market•FlexibilityGeneralembeddedprocessorarchitecturesintroducedtosatisfytheseconstraints!•DeterministicPerformance•Powerconsumption•PerformanceProcessorarchitectureNewArchitecturalParadigmFPGAASICPDesigncostTime-to-marketFlexibilityDeterminismPowerPerformanceDesigncostTime-to-marketFlexibilityDeterminismPowerPerformanceDesigncostTime-to-marketFlexibilityDeterminismPowerPerformanceNewArchitecturalParadigmFPGAASICPDeterminismPowerPerformanceDesigncostTime-to-marketFlexibilityASCPApplication-SpecificCustomizableEmbeddedProcessor–Helpspreservethebenefitsofgenerality–Alleviatesthedrawbacksofgeneral-purposeprocessorsStaticvs.DynamicOptimizationsinGeneralPurposeProcessorsApplicationApplicationInformationISAMicro-ArchitectureHardwarefordynamicresolutionStrengths:•Executioninformationavailable•Optimizesinstructioninstances•Limitedprocessingpower•Run-time“training”requiresstorage•LimitedapplicationknowledgeWeaknesses:•Globalprograminformationavailable•“Unlimited”processingpoweravailableStrengths:•ISAusedforinformationtransfer•Norun-timeinformationWeaknesses:CompilerOptimizationsArchitecturalOptimizationsDynamicallyCustomizableEmbeddedProcessorsCompilersforstaticextractionArchitecturalRuntimeincorporationASCL“shapes”theprocessorbymatchingcompilerinformationtomicroarchitectureASCLMicro-ArchitectureApplicationISAExecutionResourcesASCL–ApplicationSpecificCustomizationLogicProvidesdeterministicinformationaboutapplicationregularitiesRestrictsthedomainofpossibleapplicationbehaviorsHardwarefordynamicresolutionMicroarchitecturalCustomizationsUseofapplicationknowledgeinmicroarchitecturalmodules–Power–Performance–DeterminismReprogrammablecustomizationhardware–Post-manufacturingre-customizations–LargemanufacturingvolumesApplicationProgramRAMFU1FU2MicroarchitecturePInformationTransferandHardwareSupportLoopALoopBLoopCApplicationAppInformationASCLApplication-SpecificPSpecialregistersortablesApplicationhotspotstargetedApplicationinformationloadedintospecialhardwaretables/registers,providingreprogrammableimplementationInformationtransfereitherbysoftwareorsystemsetupUnifiedCustomizableArchitectureALURegFileFU1FU2…ASBRACBTBDecodeDataCachePartitioned/CompressedTagsLow-PowerInstr.TransformI-MemDataMemoryAunified,dynamicallycustomizableembeddedprocessorarchitectureTLBApplicationMicro-ArchitectureISAStatisticallybasedmethodsnormallyusedtoinferapplicationpropertiesPowerexpensivemicroarchitecturalcomponentsHighlysub-optimalperformanceUnpredictableexecutiontimeFU1FU2…CachesBranchPred.,etc.ApplicationKnowledgeBenefits:PreciseapplicationknowledgeApplicationMicro-ArchitectureISAStatisticallybasedmethodsnormallyusedtoinferapplicationpropertiesPreciseapplicationknowledgeusedinsteadthroughtheASCLApplicationregularitiesreadilyavailableforutilization=ScaleddownandpowerefficientuArchitecturalcomponentsDeterministicexecutiontimeachievedFU1FU2…CachesBranchPred.,etc.ASCLApplicationKnowledgeBenefits:PreciseapplicationknowledgeApplicationKnowledgeBenefits:RefinedprogrambehaviorFU1FU2…WorstcaseassumptionfortheprogramexecutionISAMicro-ArchitectureAlargesetofpotentialprograms!WorstcaseexecutionscenarioassumedingeneralpurposeprocessorP1P2P3P5P4P1P2P3P1P2P3FU1FU2…Consideringasingleprogramsegmentonly!ISAWorstcaseexecutionscenarioassumedingeneralpurposeprocessorsApplicationknowledgerefinesthedomainofallpossiblestatesRedundanthardwareactivitiesremoved=PowersavingsASCLAlargesetofpotentialprograms!P1P2P3P5P4ApplicationKnowledgeBenefits:RefinedprogrambehaviorMicro-ArchitectureConclusionsAcustomizableprocessorarchitecturedefined–In-fieldrecustomization–Highvolumesduetofixed-siliconarchitecture–AunifiedarchitecturefordiversesetsoftasksTheadaptivearchitectureprovidesflexibility,highutilization,andlowpowerforaneverincreasinganddiversesetofapplicationsExperimentallyverifiedordersofmagnitudeimprovements