Neural Processing Unit (NPU) Explained - Utmel

NPU (neural processing unit) is a specialized processor for network application packets, using a "data-driven parallel computing" architecture, ... ProductsBrandRFQTools AccountCenter Cart RFQList RegionalSettings English русский USD-$ SignInProductsInputPartNumber/KeywordHomeTheblogcategoriesintegratedcircuitNeuralProcessingUnit(NPU)ExplainedNeuralProcessingUnit(NPU)ExplainedAlan28December2021523Theembeddedneuralnetworkprocessingunit(NPU)adoptsa"data-drivenparallelcomputing"architectureandisparticularlygoodatprocessingmassivemultimediadatasuchasvideoandimages.NeuralProcessingUnitNPU,ArtificialIntelligenceAIandMachineLearningMLexplainedCatalogⅠ WhatisanNPU?Ⅱ Processormodules ofNPUⅢ NPU:thecorecarrierofcellphoneAIⅣ NPUvs. GPUⅤThecharacteristicsofdifferent processingunitsⅥ PracticalapplicationsofNPUⅦ Explanationofeachtypeofprocessingunit Ⅰ WhatisanNPU?Itisjustatthebeginningoftheexplosionofdemandforneuralnetworksandmachinelearningprocessing.TraditionalCPUs/GPUscandosimilartasks,butNPUspecificallyoptimizedforneuralnetworkscanperformmuchbetterthanCPUs/GPUs.Gradually,similarneuralnetworktaskswillbedonebydedicatedNPUunits.NPU(neuralprocessingunit)isaspecializedprocessorfornetworkapplicationpackets,usinga"data-drivenparallelcomputing"architecture,especiallygoodatprocessingmassivemultimediadatasuchasvideoandimages.NPUisalsoakindofintegratedcircuit,butdifferentfromthesinglefunctionofspecial-purposeintegratedcircuit(ASIC),networkprocessingismorecomplex,moreflexible. Generally,we canusesoftwareorhardwareinaccordancewiththecharacteristicsofnetworkcomputingspecialprogrammingtoachievethespecialpurposeofthenetwork. NPUThehighlightoftheNPUistheabilitytorunmultipleparallelthreads--TheNPUistakentoanotherlevelwithsomespecialhardware-leveloptimizations,suchasprovidingsomeeasilyaccessiblecachingsystemsforsomereallydifferentprocessingcores.Thesehigh-capacitycoresaresimplerthantypical"regular"processorsbecausetheydon'tneedtoperformmultipletypesoftasks.Thissetof"optimizations"makesNPUsmoreefficient,whichiswhysomuchR&DisbeingputintoASICs.OneoftheadvantagesofNPUsisthattheyspendmostoftheirtimefocusingonlow-precisionalgorithms,newdataflowarchitectures,orin-memorycomputingcapabilities.UnlikeGPUs,theyaremoreconcernedwiththroughputthanlatency.Ⅱ Processormodules ofNPUNPUisspeciallydesignedforIoTAItoaccelerateneuralnetworkoperationsandsolvetheproblemofinefficiencyoftraditionalchipsinneuralnetworkoperations.TheNPUprocessorincludesmodulesformultiplicationandaddition,activationfunctions,2Ddataoperations,decompression,etc.Themultiplicationandadditionmoduleisusedtocalculatematrixmultiplicationandaddition,convolution,dotproduct,andotherfunctions.Thereare64MACsinsidetheNPUand32intheSNPU. Theactivationfunctionmoduleisusedtoimplementtheactivationfunctionintheneuralnetworkbyusingthehighest12thorderparameterfitting,with6MACsinsidetheNPUand3intheSNPU.The2Ddataoperationmoduleisusedtoimplementoperationsonaplane,suchasdownsamplingandplanedatacopying,etc.Thereare1MACand1SNPUinsidetheNPU.Thedecompressionmoduleisusedtodecompresstheweighteddata.InordertosolvethecharacteristicsofsmallmemorybandwidthinIoTdevices,theweightsintheneuralnetworkarecompressedintheNPUcompiler,whichcanachievea6-10timescompressioneffectwithalmostnoimpactonaccuracy.Ⅲ NPU:thecorecarrierofcellphoneAIAsweallknow,thenormaloperationofcellphonesisinseparablefromtheSoCchip,whichisonlythesizeofafingernailcoverbuthasallthe"guts".Itsintegratedmodulesworktogethertosupporttheimplementationofcellphonefunctions. TheCPUisresponsibleforsmoothswitchingofmobileapplications,theGPUsupportsfastloadingofgamescreens,andtheNPUisspecificallyresponsiblefortheimplementationofAIcomputingandAIapplications.ItisalsonecessarytostartwithHuawei,whichisthefirstcompanytouseNPU(neural-networkprocessingunits)oncellphones,andthefirstcompanytointegrateNPUintocellphoneCPUs.In2017,HuaweilauncheditsownarchitectureNPU.Comparedwiththetraditionalscalarandvectorcomputingmode,Huawei'sself-developedarchitectureNPUuses3DCubetodoaccelerationformatrixcomputing.Asaresult,theamountofdatacalculatedperunittimeislarger,andtheAIarithmeticpowerperunitpowerconsumptionisstronger,achievinganorderofmagnitudeimprovementcomparedtotraditionalCPUsandGPUs,andachievingbetterenergyefficiency.HuaweifirstusedCambrian'sNPUfortheMate10usinganexternalapproach. A yearlater,HuaweiintegratedCambrian'sNPUIPintothe980,andanotheryearlater,HuaweiabandonedCambriantouseitsownDaVinciNPUonthe990.TheNPUinGalaxyisalsobuiltintothemobileprocessortoharnessadvancedneuralnetworksandprovideahigherlevelofvisualintelligencefortheGalaxyS20/S20+/S20UltraandZFlip.TheNPUpowersthesceneoptimizer,enhancingtheabilitytorecognizecontentinphotosandpromptingthecameratoadjustittotheidealsettingsforthesubject.ItisalsonowmoreaccuratethanthepreviousGalaxymodel.Italsoenablesthefront-facingcameratoblurthebackgroundofselfiesandcreatebokeheffects.Notonlythat,buttheNPUalsohelpsAIBixbyVisiononthedevice.Ⅳ NPUvs. GPUAlthoughGPUhastheadvantageinparallelcomputingcapability,itdoesnotworkaloneandneedstheCPU'sco-processing.TheconstructionofneuralnetworkmodelsanddatastreamsarestillcarriedoutontheCPU.Thereisalsotheproblemofhighpowerconsumptionandlargesize.Thehighertheperformance,thelargertheGPU,thehigherthepowerconsumption,andthemoreexpensiveitis, whichwillnotbeavailableforsomesmalldevicesandmobiledevices.Therefore,asmallsize,lowpowerconsumption,highcomputationalperformance,andhighcomputationalefficiencyofadedicatedchipNPUwasborn.NPUworksbysimulatinghumanneuronsandsynapsesatthecircuitlayeranddirectlyprocessinglarge-scaleneuronsandsynapseswithadeeplearninginstructionset,whereoneinstructioncompletestheprocessingofasetofneurons.ComparedwithCPUsandGPUs,NPUsintegratestorageandcomputationthroughsynapticweights,thusimprovingoperationalefficiency.CPUandGPUprocessorsneedtousethousandsofinstructionstocompletetheneuronprocessing.NPUcanbecompletedwithjustoneorafewinstructions,soithasobviousadvantagesintheprocessingefficiencyofdeeplearning.ExperimentalresultsshowthattheperformanceofNPUis118timesthatofGPUwiththesamepowerconsumption.ⅤThecharacteristicsofdifferent processingunitsCPU--70%ofthetransistorsareusedtobuildCache,andpartofthecontrolunit. Fewcomputationalunits,suitableforlogiccontroloperations.GPU--Transistorsaremostlyusedtobuildcomputationalunits,withlowcomputationalcomplexity,suitableforlarge-scaleparallelcomputing.Mainlyusedinbigdata,backendserver,imageprocessing.NPU--simulateneuronsincircuitlayer,realizestorageandcomputationintegrationbysynapticweights. Oneinstructioncompletestheprocessingofagroupofneurons,improveoperationefficiency.Mainlyusedinthecommunicationfield,bigdata,imageprocessing.FPGA--programmablelogic,highcomputationalefficiency,closertotheunderlyingIO. Logiceditablethroughredundanttransistorsandlinkages.Essentiallyinstructionless,nosharedmemoryrequired,andmorecomputationallyefficientthanCPUandGPU.Mainlyusedinsmartphones,portablemobiledevices,andautomobiles.Ⅵ PracticalapplicationsofNPUAIscenerecognitionbyNPUwhentakingphotos,andretouchingpictureswithNPUcomputing.NPUjudgeslightsourceanddarklightdetailstosynthesizesupernightscenes.RealizevoiceassistantoperationbyNPU.NPUwithGPUTurbopre-determinethenextframetoachieveearlyrenderingtoimprovethesmoothnessofthegame.NPUpre-determinesthetouchtoimprovethefollowinghandandsensitivity.NPUjudgesthedifferenceoffrontandback-endnetworkspeeddemandwithLinkTurbo.NPUjudgethegamerenderingloadtoadjusttheresolutionintelligently.GivetheNPUthepowersavingbyreducingthecomputingloadofAIduringthegame.NPUrealizesthedynamicschedulingofCPUandGPU.NPUassistsinbigdataadvertisingpush.ImplementAIintelligentwordassociationfunctionofinputmethodthroughNPU.Ⅶ ExplanationofeachtypeofprocessingunitAPU:AcceleratedProcessingUnit,AMD'sproductforacceleratedimageprocessingchip.BPU:BrainProcessingUnit,Horizon'sleadingembeddedprocessorarchitecture.CPU:CentralProcessingUnit,themainstreamproductofPCcore.DPU:DataflowProcessingUnit,anAIarchitectureproposedbyWaveComputing.FPU:FloatingProcessingUnit,thefloating-pointmoduleingeneral-purposeprocessors.GPU:GraphicsProcessingUnit,withmulti-threadedSIMDarchitecture,madeforgraphicsprocessing.HPU:HolographicsProcessingUnit,aholographiccomputingchip,anddevicefromMicrosoft.IPU:IntelligentProcessingUnit,DeepMind'sinvestmentinGraphcore'sAIprocessorproducts.MPU/MCU:Microprocessor/MicrocontrollerUnit,generallyusedforlowcomputingapplicationsofRISCcomputerarchitectureproducts,suchasARM-Mseriesprocessors.NPU:NeuralNetworkProcessingUnit,ageneraltermforanewtypeofprocessor-basedonneuralnetworkalgorithmandacceleration,suchasthediannaoseriesfromtheInstituteofComputing,ChineseAcademyofSciences/Cambrian.RPU:RadioProcessingUnitradioprocessor,acollectionofWifi/Bluetooth/FM/processorasasingleprocessorfromImaginationTechnologies.TPU:TensorProcessingUnit,adedicatedprocessorforacceleratingartificialintelligencealgorithmsfromGoogle.ThecurrentgenerationTPUforInference,thesecondgenerationfortraining.VPU:VectorProcessingUnit,adedicatedchipforacceleratingimageprocessingandartificialintelligencelaunchedbyMovidius,acompanyacquiredbyIntel.WPU:WearableProcessing,InedaSystemslaunchedwearablesystem-on-a-chipproducts,includingGPU/MIPSCPUandotherIP.XPU:BaiduandXilinx'sFPGAintelligentcloudaccelerationannouncedatHotchips2017,containing256cores.ZPU:ZylinProcessingUnit,a32-bitopensourceprocessorfromNorway'sZylin.Sharethispost FrequentlyAskedQuestions 1.WhatisneuralprocessingunitNPU?Aneuralprocessingunit(NPU)isamicroprocessorthatspecializesintheaccelerationofMachineLearningalgorithms,typicallybyoperatingonpredictivemodels.Itisalsoknownasaneuralprocessor.2.WhomakesNPU?Armtodayannouncedtwonewprocessors(oroneandahalf,dependingonhowyoulookatit).Thecompany,whichdesignsthechipsthatpowerthemajorityoftheworld'scellphonesandsmartdevices,launchedboththenewestCortex-Mprocessor(theM55)andtheArmEthos-U55microneuralprocessingunit(NPU).LeaveaCommentSUBMITRelatedArticlesHowareIntegratedCircuitsproduced?Alan17March20208598Anintegratedcircuitisaminiatureelectronicdeviceorcomponentwithspecificcircuitfunctions.Itismadethroughsemiconductormanufacturingprocessessuchasoxidation,photolithography,diffusion,epitaxy,andaluminumevaporation.ReadMore>WhatisaDigitalIntegratedCircuitandHowDoWeUseIt?Rachel18March20203932Digitalintegratedcircuitsaredigitallogiccircuitsorsystemsmadebyintegratingcomponentsandwiresonthesamesemiconductorchip.Accordingtothenumberofincludedgatecircuitsorcomponents,digitalintegratedcircuitscanbedividedintoSmallScaleIntegrated(SSI)circuits,MediumScaleIntegrated(MSI)circuits,LargeScaleIntegration(LSI)circuits,VeryLargeScaleIntegrated(VLSI)circuits,andUltraLargeScaleIntegration(ULSI)circuits.ReadMore>IntroductiontoPhotonicIntegratedCircuitsandPICTechnologyRachel09April20203618Theconceptofthephotonicintegratedcircuitissimilartotheconceptofelectronicintegratedcircuits.However,electronicintegratedcircuitsintegratetransistors,capacitors,resistors,andotherelectronicdevices,whilePICintegratesvariousopticaldevicesoroptoelectronicdevices,suchaslasers,electro-opticalmodulators,photodetectors,opticalattenuators,opticalmultiplexers/demultiplexers,andopticalamplifiers.ReadMore>Types,Structure,andPackagesofIntegratedCircuitsAlan30April20202997Thetypeofintegratedcircuitcanbedividedaccordingtoitsuse,function,structure,manufacturingprocess,integration,andconductivitytype.Thestructureoftheintegratedcircuitcanbelearnedfromlevelsfromtoptobottom:systemlevel-modulelevel-registertransferlevel(RTL)-gatelevel-transistorlevel.ThepackagingformsofintegratedcircuitsincludeSOP,PGA,BGA,DIP.ReadMore>AnOverviewofLinearIntegratedCircuitsRachel08May20202775Alinearintegratedcircuitisanintegratedcircuitbasedonanamplifier.Themeaningof"linear"usedtoindicatethattheresponserelationshipoftheamplifiertotheinputsignalisusuallylinear.Inthisessay,we’llfirstintroducethedevelopmenthistoryoflinearintegratedcircuit.Andthentheproductionprocessandthehigh-frequencytechnologyarediscussedindetail.Thefollowingtwopartsmainlytalkaboutthelinearintegratedcircuittypesandforms.Andthefinalchapteranalyzesthetwotypicalbipolarlinearintegratedcircuit一Lm386and34063.ReadMore>SubscribetoUtmel!SUBSCRIBEPopularPostsHowmanyTransistorsinaCPU?Whatisavariablecapacitor?Varistor:Definition,Function,WorkingandTestingPhotoresistorBasics:Types,PrinciplesandApplicationsWorkingPrincipleandCharacteristicsofZenerdiodesStructureandWorkingPrincipleofFieldEffectTransistorsHowareIntegratedCircuitsproduced?WhatisaMemoryController?WhatisaVariableResistor?WorkingPrinciplesandApplicationsofPressureSensorsRelatedParts More MX25L12835FZ2I-10GMacronixPackage:8-WDFNExposedPadICFLASH128MSPI104MHZ8WSONAT25DF321A-SH-BAdestoTechnologiesPackage:8-SOIC(0.209,5.30mmWidth)NORFlashSerial(SPI,DualSPI)3.3V32M-bit4Mx85ns8-PinSOICEIAJBulkMX29LV320EBTI-70GMacronixPackage:48-TFSOP(0.724,18.40mmWidth)ICFLASH32MPARALLEL48TSOPMX29GL640ELT2I-70GMacronixPackage:56-TFSOP(0.724,18.40mmWidth)ICFLASH64MPARALLEL56TSOPMX25L25635FZ2I-10GMacronixPackage:8-WDFNExposedPadICFLASH256MSPI104MHZ8WSONMX29GL256FHXFI-90QMacronixPackage:64-LBGA,CSPBGAICFLASH256MPARALLEL64LFBGAMX66L51235FZ2I-10GMacronixPackage:8-WDFNExposedPadICFLASH512MSPI104MHZ8WSONM48Z12-150PC1STMicroelectronicsPackage:24-DIPModule(0.600,15.24mm)STMICROELECTRONICSM48Z12-150PC1IC,SRAM,16KB,5VTop AccountCenter(0Items) Opps.It'sempty.10,000+DailyOrderQuantity5000,000+AlternativeParts2800+WorldwideManufacturers32,000㎡In-stockWarehouse Utmelusescookiestohelpdeliverabetteronlineexperience.Youcanseewhatcookiesweserveandhowtosetyourpreferencesinour CookiesPolicy, ifyouagreeonouruseofcookiespleaseclickcontinue.Whenbrowsingandusingourwebsite,Utmelalsocollects,storesand/orprocessespersonaldata,pleasereadour Term&Condition and PrivacyPolicy tofindoutmore. CONTINUE