What Is An xPU? - Semiconductor Engineering
文章推薦指數: 80 %
Almost every day there is an announcement about a new processor architecture, and it is given a three-letter acronym — TPU, IPU, NPU. Home Systems&Design LowPower-HighPerformance Manufacturing,Packaging&Materials Test,Measurement&Analytics Auto,Security&PervasiveComputing SpecialReports Videos Jobs KnowledgeCenter TechnicalPapers Home'; AI/ML/DLArchitecturesAutomotiveCommunication/DataMovementDesign&VerificationLithographyManufacturingMaterialsMemoryOptoelectronicsPackagingPower&PerformanceQuantumSecurityTest&AnalyticsTransistors Events&Webinars Events Webinars Research&Startups IndustryResearch StartupCorner MENU Home SpecialReports Systems&Design LowPower-HighPerformance Manufacturing,Packaging&Materials Test,Measurement&Analytics Auto,Security&PervasiveComputing KnowledgeCenter Videos StartupCorner Jobs TechnicalPapers Events Webinars IndustryResearch SpecialReports Home> LowPower-HighPerformance> WhatIsAnxPU? LowPower-HighPerformance WhatIsAnxPU? Almosteveryletterofthealphabethasbeenusedtodescribeaprocessorarchitecture,butunderthehoodtheyalllookverysimilar. November11th, 2021- By:BrianBailey Almosteverydaythereisanannouncementaboutanewprocessorarchitecture,anditisgivenathree-letteracronym—TPU,IPU,NPU.Butwhatreallydistinguishesthem?Aretherereallythatmanyuniqueprocessorarchitectures,orissomethingelsehappening? In2018,JohnL.HennessyandDavidA.PattersondeliveredtheTuringlectureentitled,“ANewGoldenAgeforComputerArchitecture.”WhattheyconcentratedonwastheCPUandthewaythatithadevolved,butthatisonlyasmallpartofthetotalequation.“MostofthesethingsarenotreallyaprocessorinthesenseofbeingaCPU,”saysMichaelFrank,fellowandsystemarchitectatArterisIP.“They’remorelikeaGPU,anacceleratorforaspecialworkload,andthereisquiteabitofdiversitywithinthem.Machinelearningisaclassofprocessors,andyoujustcallthemallmachinelearningaccelerators,yetthereisalargevarietyofthepartoftheprocessingtheyaccelerate.” Theessenceofaprocessorcanbeboileddowntothreethings.“Attheendoftheday,itreallydoescomebacktotheinstructionsetarchitecture(ISA),”saysManuelUhm,directorforsiliconmarketingatXilinx.“Thatdefineswhatyou’retryingtodo.ThenyouhavetheI/Oandthememory,whichsupporttheISAandwhatit’stryingtoaccomplish.It’sgoingtobeareallyinterestingtimegoingforward,becausewearegoingseealotmoreinnovationandchangethanwe’veseeninthelasttwo-orthree-plusdecades.” Manyofthenewarchitecturesarenotsingleprocessors.“Whatweareseeingisacombinationofdifferenttypesofprocessors,orprogrammableengines,thatliveinthesameSoCorinthesamesystem,”saysPierre-XavierThomas,groupdirectorfortechnicalandstrategicmarketingatCadence.“Thereisdispatchingofthesoftwaretaskstodifferenthardwareorflexibleprogrammableengines.AlloftheprocessorsmayshareacommonAPI,buttheexecutiondomainisgoingtobedifferent.Thisisreallywhereyouwillseedifferenttypesofprocessingwithdifferenttypeofcharacteristics.” Therealityisthatmuchofthenamingismarketing.“Thekeythingisthatpeopleareusingthesenamesandacronymsfortwodifferentpurposes,”saysSimonDavidmann,CEOforImperasSoftware.“Oneisforexplainingthearchitectureofaprocessor,likeSIMD(singleinstructionmultipledata).Theotherdefinestheapplicationsegmentthatitisaddressing.Soitcandefineeithertheprocessorarchitecture,orabrandnamelikeTensorProcessingUnit(TPU).Theyareputtinganametotheirheterogeneousorhomogeneousarchitecture,whichisnotasingleprocessor.” Alittlehistory Thingsweremuchsimpler40yearsago.Therewasthecentralprocessingunit(CPU),andwhilethereweremanyvariantsofit,theywereallfundamentallyvonNeumanarchitecture,Turingcompleteprocessors.Eachhaddifferentinstructionsetsthatmadethemmoreefficientforcertaintasksandtherewasplentyofdiscussionabouttherelativemeritofcomplexinstructionset(CISC)versusreducedinstructionset(RISC). TheemergenceofRISC-VbroughtalotofattentiontotheISA.“PeoplewanttounderstandtheISAbecauseitistheISAthatdefineshowoptimizedtheprocessorisforadefinedtask,”saysXilinx’sUhm.“TheycanlookattheISAandstartcountingcycles.IfoneISAhasanativeinstructionandoperatesatonegigahertz,IcancomparethattoanotherprocessorISAwherethesamefunctionmayrequiretwoinstructions,buttheprocessorrunsat1.5gigahertz.Whichonegetsmefurtherahead?Theydothemathfortheimportantfunctionality.” CPUshavebeenpackagedinmanyways,sometimesputtingIOormemoryintothesamepackageandtheywerecalledmicro-controllerunits(MCU). Whenmodemsbecamefashionable,digitalsignalprocessors(DSP)emerged,andtheyweredifferentbecausetheyusedtheHarvardarchitecture.Thatseparatedtheinstructionbusfromthedatabus.SomeofthemalsoimplementedSIMDarchitecturesthatmadedatacrunchingmoreefficient. Theseparationofinstructionsanddatawasdonetoincreasethroughputrates,eventhoughitrestrictedsomefringeprogrammingthatcouldbedone,suchasself-writingprograms.“Often,itisnotcomputethatistheboundarycondition,”saysUhm.“ItisincreasinglytheI/Oormemory.Theindustryswitchedfromjackingupcompute,tomakingsurethatthere’senoughdatatokeepthecomputecrunchingandmaintainperformance.” Whensingleprocessorsstoppedbecomingfaster,multipleprocessorswerelinkedtogether,oftensharingmemoryandmaintainingthenotionthateachprocessor,andthetotalclusterofprocessors,remainTuringcomplete.Itdidn’tmatterwhichcoreanypieceofaprogramwasexecutedon.Theresultwasthesame. Thenextmajordevelopmentwasthegraphicsprocessingunit(GPU),andthisbrokethemoldbecauseeachprocessingelementorpipelinehaditsownmemorythatwasnotaddressableoutsideoftheprocessor.Becausethememorywasfinite,itmeantthatitcouldnotperformanyarbitraryprocessingtask,onlytheonesthatcouldfitintheprovidedmemoryspace. “GPUsareverycapableprocessorsforcertaintypeoffunctions,buttheyhaveextremelylongpipelines,”notesUhm.“ThosepipelineskeeptheGPUunitscrunchingondata,butatsomepoint,ifyouhavetoflushthepipeline,that’sahugehit.Thereisasignificantamountoflatencyandnon-determinismbuiltintothesystem.” Whilemanyotheracceleratorshavebeendefined,theGPU—andlaterthegeneral-purposeGPU(GPGPU)—definedaprogrammingparadigmandsoftwarestackthatmadethemmoreapproachablethanacceleratorsofthepast.“Overtheyears,certainjobshavebeenspecialized,”saysImperas’Davidmann.“TherewastheCPUforsequentialprograms.Therewasthegraphicsprocessor,whichfocusedonmanipulationofdataforascreenandintroducedustoahighlyparallelworld.Taskswereperformedusinglotsoflittleprocessingelements.Andnowtherearemachinelearningtasks.” Whatotherconstructionrulesaretheretobebrokenthatcanexplainallofthenewarchitectures?Inthepastprocessorarrayswereoftenconnectedthroughmemory,orafixednetworktopology,suchasmeshortoroid.Whathasemergedmorerecentlyistheincorporationofanetworkonchip(NoC)thatenablesdistributed,heterogenousprocessorstocommunicateinamoreflexiblemanner.Inthefuture,theyalsomayenablecommunicationswithoutusingmemory. “Atthispoint,NoCsonlycarrydata,”saysArteris’Frank.“Inthefuture,theNoCcouldexpandintootherareaswherecommunicationbetweenacceleratorsgoesbeyonddata.Itcouldsendcommands,itcouldsendnotifications,etc.Thecommunicationneedsofanacceleratorarrayor,seaofaccelerators,mightbedifferentthanthecommunicationneedsof,forexample,CPUsorastandardSoC.Butnetworkonachipdoesnotconstrainyoutojustasubset.Youcanoptimizeandimproveperformancebysupportingspecialcommunicationneedsofaccelerators.” Implementationarchitecture Onewaythatprocessorsdifferentiateisbyoptimizingforaparticularoperatingenvironment.Forexample,softwaremayruninthecloud,butyoumayalsoexecutethesamesoftwareonatinyIoTdevice.Theimplementationarchitecturewillbeverydifferentandachievedifferentoperatingpointsintermsofperformance,power,cost,ortheabilitytooperateunderextremeconditions. “Someapplicationsweretargetedforthecloud,andnowwe’rebringingthemclosertotheedge,”saysCadence’sThomas.“Thismaybebecauseoflatencyrequirements,orforenergyorpowerdissipation,andthatwouldrequireadifferenttypeofarchitecture.Youmaywanttohaveexactlythesamesoftwarestacktobeabletoruninbothlocations.Thecloudneedstoprovideflexibilitybecauseitwillbereceivingdifferenttypesofapplicationsandhastobeabletoaggregateanumberofusers.Thisrequiresthehardwareontheservertobeapplication-specificcapable,butonesizedoesnotfitall.” MLaddsitsownrequirements.“Whenbuildingintelligentsystemswithneuralnetworksandmachinelearning,youneedtoprogramnewnetworksandmapthistohardware,usingsoftwareframeworksandacommonsoftwarestack,”addsThomas.“YoucanthenadaptthesoftwareapplicationtotherighthardwarefromaPPAstandpoint.Thisdrivestheneedfordifferenttypesofprocessingandprocessorstobeabletoaddresstheseneedsatthehardwarelayer.” Thoseneedsaredefinedbytheapplication.“Onecompanyhascreatedaprocessorforgraphoperations,”saysFrank.“Theyoptimizeandacceleratehowtofollowgraphs,anddooperationssuchasreorderingofgraphs.Thereareothersthatmostlyacceleratethebruteforcepartofmachinelearning,whichismatrix-matrixmultiplies.Memoryaccessisaparticularproblemforeacharchitecture,becausewhenyoubuildanaccelerator,themostimportantgoalistokeepitbusy.YouhavetogetasmuchdatathroughtotheALUsasitcanconsumeandproduce.” Manyoftheseapplicationshaveanumberofthingsincommon.“Theyallhavesomelocalmemory,theyhaveanetworkonchiptocommunicatethingsaround,andeachprocessor,whichexecutesasoftwarealgorithm,iscrunchingonasmallchunkofdata,”saysDavidmann.“ThosejobsarescheduledbyanOSwhichrunsonamoreconventionalCPU.” Thetrickybitforhardwaredesignersispredictingwhattasksitwillbeaskedtoperform.“Althoughyou’regoingtohavesimilartypesofoperationinsomeofthelayers,peoplearelookingatdifferentiationinthelayers,”saysThomas.“Tobeabletoprocesstheneuralnetworkrequiredseveraltypesofprocessingcapabilities.Itmeansthatyouneedtobeabletoprocessacertainwayforonepartoftheneuralnetwork,andthenanothertypeofoperationsmayberequiredtoprocessanotherlayer.Thedatamovementandtheamountofdataisalsochanginglayerafterlayer.” Thisdifferentiationcangobeyondthedatamovement.“Forgenomesequencing,youneedtodocertainprocessing,”saysFrank.“Butyoucannotaccelerateeverythingwithasingletypeofaccelerator.Youhavetobuildacompletesetofdifferentacceleratorsforthepipeline.TheCPUsbecometheguardianthatshepherdtheexecutionflow.Itsetsthingsup,doestheDMA,providesthedecision-makingprocessinbetween.Thereisawholearchitecturetasktounderstandandanalyzealgorithmsanddefinehowyouwanttooptimizetheprocessingofthem.” Partofthatprocessrequirespartitioning.“Thereisnosingleprocessortypethat’soptimizedforeverysingleprocessortask—notFPGAs,notCPUs,notGPUs,notDSPs,”saysUhm.“Wecreatedaseriesofdevicesthatcontainallofthose,butthehardpartonthecustomersideisthattheyhavetoprovidetheintelligencetodeterminewhichpartsofthisentiresystemaregoingtobetargetedattheprocessors,orattheprogrammablelogic,orattheAIengines.Everyonewantsthatauto-magicaltool,atoolthatcaninstantlydecidetoputthisontheCPU,putthatontheFPGA,putthatontheGPU.Thattooldoesnotexisttoday.” Still,therealwayswillbearolefortheCPU.“CPUsareneededtoexecutetheirregularpartoftheprogram,”saysFrank.“ThegeneralprogrammabilityoftheCPUhasitsadvantages.Itjustdoesn’tworkwellifyouhavespecializeddatastructuresormathematicaloperations.ACPUisageneralprocessor,anditisnotoptimizedforanything.It’sgoodatnothing.” Changingabstraction Inthepast,thehardware/softwareboundarywasdefinedbytheISA,andthatmemorywascontiguouslyaddressable.Whenmultipleprocessorsexisted,theyweregenerallymemory-coherent. “Coherenceisacontract,”saysFrank.“Itisacontractbetweenagentsthatsays,‘IpromiseyouthatIwillalwaysprovidethelatestdatatoyou.’Coherencebetweenequalpeersisveryimportantandwillnotgoaway.Butyoucouldimaginethatinadataflowengine,coherenceislessimportantbecauseyou’reshippingthedatathatismovingontheedge,directlyfromoneacceleratortotheother.Ifyoupartitionthedataset,coherencegetsinthewaybecauseitcostsyouextracycles.Youhavetolookthingsup.Youhavetoprovidetheupdateinformation.” Thatcallsfordifferentmemoryarchitectures.“Youhavetothinkaboutthememorystructurebecauseyouonlyhavesomuchtightlycoupledmemory,”saysUhm.“Youcouldaccessadjacentmemory,butyouquicklyrunoutofadjacenciestobeabletodothatinatimelyfashion.Thathastobecomprehendedinthedesign.Asthetoolsmature,moreofthatwillstarttobecomeunderstoodbythetools.Todayitisdonebyhumanintelligence,bybeingabletounderstandthearchitectureandapplyit.” Therealsoisaneedforhigherlevelsofabstraction.“Thereareframeworkswhereyoucanmap,orcompile,knownnetworksontotargethardware,”saysThomas.“Youhaveasetoflow-levelkernels,orAPIs,thatwillbeusedinthesoftwarestack,andtheneventuallyusedbythemapperoftheneuralnetwork.Underneath,youmayhavedifferenttypesofhardware,dependingonwhatyouwanttoachieve,dependingonyourproductdetails.Itimplementsthesamefunctionality,butnotwiththesamehardware,notonthesamePPAtradeoff.” Thatputsalotofpressureonthosecompilers.“Themainquestionishowdoyouprogramacceleratorsinthefuture?”asksFrank.“DoyouimplementhardwiredenginesthatarejuststrungtogetherlikethefirstgenerationofGPUs?Ordoyoubuildlittleprogrammableenginesthathavetheirowninstructionset?Andnowyouhavetogoandprogramthesethingsindividuallyandconnecteachoftheseengines,executingtasks,withadataflow.Oneprocessorhassomesubsetofthetotalinstructionset,anotheronehasadifferentsubset,andtheywillallsharesomeoverlappingpartforthecontrolflow.Youmighthavesomethathaveslightlydifferentaccelerationcapabilities.Thecompilers,orthelibrariesthatknowaboutit,mapaccordingly.” Conclusion Thearchitectureofprocessorsisnotchanging.Theystillabidebythesamechoicesthathaveexistedforthepast40years.Whatischangingisthewayinwhichchipsarebeingconstructed.Theynowcontainlargenumbersofheterogeneousprocessorsthathavememoryandcommunicationsoptimizedforasubsetofapplicationtasks.Eachchiphasmadedifferentchoicesabouttheprocessorcapabilitiesandwhattheyareoptimizedfor,abouttherequireddatathroughput,andaboutthedataflowsthattypicallywillbeseen. Everyhardwareproviderwantstodifferentiateitschipfromtheothers,butthat’saloteasiertodothatbybrandingthanbytalkingaboutthetechnicaldetailsoftheinternals.Andsotheygiveitanameandcallitthefirst,thefastest,thelargest,andtieittoaparticulartypeofapplicationproblem.Thethreeletteracronymshavebecomeapplicationtasknames,buttheydonotdefinethehardwarearchitecture. Related ChallengesForNewAIProcessorArchitectures GettinganAIseatinthedatacenterisattractingalotofinvestment,buttherearehugeheadwinds. TenLessonsFromThreeGenerationsShapedGoogle’sTPUv4i EvolutionofGoogle’sTPUv4i NewArchitectures,MuchFasterChips Massiveinnovationtodriveordersofmagnitudeimprovementsinperformance. ImprovingMedicalImageProcessingWithAI Faster,smarterimagingopensdoorstoeverythingfrom4Dmodelingandhigherresolutionwithlessnoise. Tags:architecturesArterisIPCadenceCadenceDesignSystemsCPUDSPFPGAGPUHarvardarchitectureImperasSoftwareIPUMCUnetworkonchipNPUprocessorsSIMDTPUTuringcompleteXilinx BrianBailey (allposts) BrianBaileyisTechnologyEditor/EDAforSemiconductorEngineering. LeaveaReplyCancelreplyComment*Name*(Note:Thisnamewillbedisplayedpublicly) Email*(Thiswillnotbedisplayedpublicly) Δ KnowledgeCentersBlogs DigitalSignalProcessor(DSP) PublishedonNovember13,2019 CentralProcessingUnit(CPU) PublishedonOctober18,2019 Processors PublishedonJuly25,2019 GraphicsProcessingUnit(GPU) PublishedonJuly3,2019 Architectures PublishedonJuly25,2017 TechnicalPapers BenchmarkingMemory-CentricComputingSystems:AnalysisofRealProcessing-in-MemoryHardwareMarch10,2022byTechnicalPaperLinkQubiC:AnOpen-SourceFPGA-BasedControlandMeasurementSystemforSuperconductingQuantumInformationProcessorsMarch9,2022byTechnicalPaperLinkPinpointingtheDominantComponentofContactResistancetoAtomicallyThinSemiconductorsMarch8,2022byTechnicalPaperLinkSparseP:TowardsEfficientSparseMatrixVectorMultiplicationonRealProcessing-In-MemorySystemsMarch8,2022byTechnicalPaperLinkZero-BiasPower-DetectorCircuitsbasedonMoS2Field-EffectTransistorsonWafer-ScaleFlexibleSubstratesMarch8,2022byTechnicalPaperLink TrendingArticles WhyRISC-VIsSucceeding InterestinthisparticularISAisexpanding,butthegrowthofotheropen-sourcehardwareislesscertain. byBrianBailey TransistorsReachTippingPointAt3nm Nanosheetsarelikeliestoptionthroughoutthisdecade,withCFETsandotherexoticstructurespossibleafterthat. byMarkLaPedus MachineLearningShowingUpAsSiliconIP Itwon’treplaceMLchips,butitcouldbroadenthemarket. byBryonMoyer TechnologyAdvances,ShortagesSeenForWireBonders Oldtechstillaccountsfor75%ofallpackages,andlikelywillcontinuetoplayimportantroleasequipmentimproves. byMarkLaPedus Silicon-basedPowerSemisFaceChallenges SiCandGaNaregainingtraction,butsiliconismakingprogress,too. byMarkLaPedus KnowledgeCentersEntities,peopleandtechnologiesexplored LearnMore RelatedArticles AMinimalRISC-V IsthereroomforanevensmallerversionofaRISC-Vprocessorthatcouldreplace8-bitmicrocontrollers? byBrianBailey IsProgrammableOverheadWorthTheCost? Howmuchdowepayforasystemtobeprogrammable?Itdependsuponwhoyouask. byBrianBailey RevvingUpSiCAndGaN Technologiesarebeingdesignedintomoresystemsasdefectivitydropsandreliabilityincreases. byAnnStefforaMutschler SOT-MRAMToChallengeSRAM Spin-orbittorquememoryaddsenduranceandfasterwritespeeds,butdisplacingexistingmemoriesisstillnoteasy. byBryonMoyer ImprovingEnergyAndPowerEfficiencyInTheDataCenter Optimizingenergyandpowerefficiencyinserverchipsandsoftwareisamultifacetedchallengewithmanymovingparts. byAnnStefforaMutschler ImprovingPPAInComplexDesignsWithAI Researchshowssignificantimprovementintimetomarketandoptimizationofkeymetrics. byJohnKoon InnovationsInSensorTechnology Bettermanagementofdata,increasedaccuracy,andlowerpowerapproachestopthelist. byJohnKoon WhyDataCenterPowerWillNeverComeDown Efficiencyisimprovingsignificantly,buttheamountofdataisgrowingfaster. byBryonMoyer Sponsors Advertisewithus Advertisewithus Advertisewithus NewsletterSignup PopularTags2.5D 5G 7nm AI ANSYS Apple AppliedMaterials ARM Atrenta automotive business Cadence EDA eSilicon EUV finFETs GlobalFoundries IBM IMEC Intel IoT IP LamResearch machinelearning memory Mentor MentorGraphics Moore'sLaw Nvidia NXP OneSpinSolutions Qualcomm Rambus Samsung security SEMI Siemens SiemensEDA software Sonics Synopsys TSMC UMC verification Xilinx RecentCommentsSteveonConstraintsOnTheElectricityGridwilliam(Bill)JAtkinsononTechnicalPapers:Organized,Timely,AndRelevantAndydengonWhyBanksShouldBeMoreWorriedAboutSecurityPrinceJonUnderstandingMemoryEfiRotemonChallengesGrowForFindingChipDefectsSAmerDiabonABreakthroughInSiliconBring-UpBrianBaileyonWhyRISC-VIsSucceedingJasononTestEngineersInVeryShortSupplysnedunurionWillSteeringWheelsEverDisappear?RupertBainesonWhichProcessorIsBest?LawrenceonWhyRISC-VIsSucceedingdevdutton2DSemiconductorsMakeProgress,ButSlowlysolidproesonHowToSolveAutomotiveElectricalDesignChallengesToGetToMarketFasterPaulLueonWhyRISC-VIsSucceedingSiddharthaNathonDoesEDASellFear?DeanFreedonDoesEDASellFear?HarryChenonDoesEDASellFear?EdgarAnckeronWhy450mmwafers?DaleonAutomotiveFunctionalSafetyComplianceInEDAToolsAndIPKvsonWhyRISC-VIsSucceedingRajeevVadjikaronTransistorsReachTippingPointAt3nmMarkLaPedusonNext-Gen3DChip/PackagingRaceBeginsArnaudPHELIPOTonWhatCausesSemiconductorAging?TanjBennettonThermalManagementImplicationsForHeterogeneousIntegratedPackagingMarkLaPedusonTechnologyAdvances,ShortagesSeenForWireBondersJanHoppeonWhatCausesSemiconductorAging?RayBarrettonUnknownsDrivingUpTheCostOfAutoICReliabilityEdSperlingonWhyDataCenterPowerWillNeverComeDownMartinBuehringonUnknownsDrivingUpTheCostOfAutoICReliabilityConradChompffonWhyDataCenterPowerWillNeverComeDownJohnR.ThomeonPreparingFor3D-ICsAlexVoronelonTechnologyAdvances,ShortagesSeenForWireBondersCraigFranklinonFutureChallengesForAdvancedPackagingMy2centsonSpreadsheets:StillValuable,ButMoreLimitedLuonWhatCausesSemiconductorAging?HabibHichrionWhyDataCenterPowerWillNeverComeDownDavidLearyonWhatCausesSemiconductorAging?BesselFunconWhatCausesSemiconductorAging?AllenRasafaronCriticalMoves:AdvancedLogicDevicesAndCISBenefitFromApplicationsUsingIRCDMetrologyJacekTyminskionPowerOptimization:What’sNext?AndrewonAMinimalRISC-VNicolasBarononNext-Gen3DChip/PackagingRaceBeginsMarkDLaPedusonWeekInReview:Manufacturing,TestSantoshPrajapationDataSecurityChallengesInAutomotivedevduttonMLFocusShiftingTowardSoftwareAllenRasafaronWeekInReview:Manufacturing,TestMarkLaPedusonNext-Gen3DChip/PackagingRaceBeginsMarkLaPedusonIndustryPushesForFabToolSecurityStandardsPaulLueonNext-Gen3DChip/PackagingRaceBeginsBrianBaileyonAMinimalRISC-VBrianBaileyonAMinimalRISC-VdevduttonNext-Gen3DChip/PackagingRaceBeginsVikramChouguleonFromSandToWafersTheodoreWilsononEthicalCoverageNIRVANAonWaitingForChipletStandardsRajeevVadjikaronPhotomaskChallengesAt3nmAndBeyondMarkLaPedusonChasingAfterCarbonNanotubeFETsMarkLaPedusonEndInSightForChipShortages?MarkLaPeduson200mmShortagesMayPersistForYearsReynoldFunkonWillSteeringWheelsEverDisappear?SaumyaThackeronTheHighButOftenUnnecessaryCostOfCoherenceLarryDannenbergon200mmShortagesMayPersistForYearsArnieonEndInSightForChipShortages?S.MeirowskyonAMinimalRISC-VWillonAMinimalRISC-VTMS-EEonAMinimalRISC-VAllenRasafaronManufacturingBits:Jan.3BARRYDENNISonChasingAfterCarbonNanotubeFETsJohnSanthoffonPiecingTogetherChipletsTinaFrancisonBuildingMultipurposeSystemsWithDynamicFunctionExchangeAllenRasafaronPreventingFailuresBeforeTheyOccurTomSomyakonInspectingAndTestingGaNPowerSemisDr.MJZarabi,PhD.onManufacturingBits:Jan.10JanHoppeonManufacturingBits:Jan.10KhachikSahakyan(GrovfInc.)onTheHighButOftenUnnecessaryCostOfCoherenceHugoPristauzonNextStepsForPanel-LevelPackagingAllenRasafaronManufacturingBits:Dec.28BernardonTheHighButOftenUnnecessaryCostOfCoherenceEranWeisonReliabilityConcernsShiftLeftIntoChipDesignJnanadarshanNayakonTooMuchFabAndTestData,LowUtilizationAllenRasafaronBigPaybackForCombiningDifferentTypesOfFabDatareyhanonAdaptiveNN-BasedRootCauseAnalysisinVolumeDiagnosisforYieldImprovementKurtShuleronProductLifecycleManagementForSemiconductorsOnriJayBenallyonGearingUpForHigh-NAEUVAnsweringonInsideIntel’sAmbitiousRoadmapDr.DevGuptaonChinaAcceleratesFoundry,PowerSemiEffortsJanHoppeonChinaAcceleratesFoundry,PowerSemiEffortsBorisPetrovonChinaAcceleratesFoundry,PowerSemiEffortsLullabyonChinaAcceleratesFoundry,PowerSemiEffortsRafaelRoquesonApple’sFirstGaNChargerMarkDLaPedusonWhat’sNextForTransistorsAndChipletsMarkDLaPedusonWhat’sNextForTransistorsAndChipletsAhsanIslamonCopingWithParallelTestSite-to-SiteVariationWesNeumeieronChipmakersGettingSeriousAboutIntegratedPhotonicsguestonWhat’sNextForTransistorsAndChipletsNewtononWhat’sAfterFinFETs?petegasperinionElasticityWithoutCompromiseJacquelineArsivaudonChangingServerArchitecturesInTheDataCenterFrederikGrønSchackonNeuromorphicelectronicsbasedoncopyingandpastingthebrainJanHoppeonMoreErrors,MoreCorrectioninMemoriesOlivierSentieysonHigh-LevelSynthesisForRISC-VSimononHigh-LevelSynthesisForRISC-VTomKunichonFunctionalSafetyAcrossAnalogAndDigitalDomainsMikeDonTheBenefitsOfAntifuseOTPDanielPaneonDealingWithMarketShiftsTanjBonWhat’sNextForTransistorsAndChipletsYasunoriYamaguchionGrowingChallengesWithWaferBumpInspectionAllenRasafaronGearingUpForHigh-NAEUVJanHoppeonWhat’sNextForTransistorsAndChipletsCraigFranklinonFan-OutAndPackagingChallengesSteveTonSixThingsWeMightNeedForPervasiveComputingASAonPCBAndICTechnologiesMeetInTheMiddleAllenRasafaronEUV’sUncertainFutureAt3nmAndBelowDavidLearyonHowChipsAgeAnandChamarthyonGrapheneandtwo-dimensionalmaterialsforsilicontechnologyAbbasonTheVerificationMindsetPeterAJvanderMadeonHowMuchPowerWillAIChipsUse?MaxTurneronWillAutomotiveEthernetWin?SUHAIMISELIMANonWeekInReview:Manufacturing,TestEriconAdvantagesOfLPDDR5:ANewClockingScheme WeekInReview:Manufacturing,TestMarkLaPedusWeekInReview:Design,LowPowerJesseAllen About Aboutus Contactus AdvertisingonSemiEng NewsletterSignUp Navigation Homepage SpecialReports Systems&Design LowPower-HighPerf Manufacturing,Packaging&Materials Test,Measurement&Analytics Auto,Security&PervasiveComputing Videos Jobs TechnicalPapers Events Webinars KnowledgeCenters StartupCorner Bus&MarketingStrategies ConnectWithUs Facebook Twitter@semiEngineering LinkedIn YouTube Copyright©2013-2022SMG | TermsofService | PrivacyPolicy Thissiteusescookies.Bycontinuingtouseourwebsite,youconsenttoourCookiesPolicyACCEPTManageconsent Close PrivacyOverview Thiswebsiteusescookiestoimproveyourexperiencewhileyounavigatethroughthewebsite.Thecookiesthatarecategorizedasnecessaryarestoredonyourbrowserastheyareessentialfortheworkingofbasicfunctionalitiesofthewebsite.Wealsousethird-partycookiesthathelpusanalyzeandunderstandhowyouusethiswebsite.Wedonotsellanypersonalinformation. Bycontinuingtouseourwebsite,youconsenttoourPrivacyPolicy.Ifyouaccessotherwebsitesusingthelinksprovided,pleasebeawaretheymayhavetheirownprivacypolicies,andwedonotacceptanyresponsibilityorliabilityforthesepoliciesorforanypersonaldatawhichmaybecollectedthroughthesesites.Pleasecheckthesepoliciesbeforeyousubmitanypersonalinformationtothesesites. Necessary Necessary AlwaysEnabled Necessarycookiesareabsolutelyessentialforthewebsitetofunctionproperly.Thiscategoryonlyincludescookiesthatensuresbasicfunctionalitiesandsecurityfeaturesofthewebsite.Thesecookiesdonotstoreanypersonalinformation. Non-necessary Non-necessary Anycookiesthatmaynotbeparticularlynecessaryforthewebsitetofunctionandisusedspecificallytocollectuserpersonaldataviaanalytics,ads,otherembeddedcontentsaretermedasnon-necessarycookies.Itismandatorytoprocureuserconsentpriortorunningthesecookiesonyourwebsite. SAVE&ACCEPT
延伸文章資訊
- 1英特爾發表資料中心用Intel Server GPU,藉OneAPI 加速XPU ...
... 世代,焦點逐漸從單一處理器本身,移往橫跨CPU、GPU、FPGA 以及其它加速器的混合式架構。處理器龍頭英特爾(intel) 將這些統合描繪成「XPU」願景。
- 2琳琅满目的XPU,到底是什么? | 雷峰网
从CPU 及其发展方式的角度来看,这些“XPU”中的大部分都不是真正的处理器。机器学习加速器是一类处理器,但它们用来加速的处理部分却多种多样。
- 3聽說過CPU、GPU 琳琅滿目的XPU到底是什麼?
從CPU 及其發展方式的角度來看,這些「XPU」中的大部分都不是真正的處理器。機器學習加速器是一類處理器,但它們用來加速的處理部分卻多種多樣。
- 4使用XPU 為每個應用程式搭配最佳架構 - Intel
理想的解決方案將應用程式與其最佳硬體架構搭配,以滿足您的需求,實現合適的效能、延遲與功耗。CPU、GPU、FGPA、VPU、ASICs,甚至是神經運算等架構,都是Intel XPU 策略的 ...
- 5看懂全球AI晶元,詳解「xPU」,這可能是至今最全的一份知識點
APU是AMD的一個處理器品牌。AMD在一顆晶元上集成傳統CPU和圖形處理器GPU,這樣主板上將不再需要北橋,任務可以靈活地 ...