人工智能：引領未來深度學習機器

PleasereadtheImportantNoticeandWarningsattheendofthisdocument

v1.0

WhitePaper

04-2017

Artificialintelligence:poweringthedeep-learningmachinesoftomorrow

Deeplearningneuralnetworksdemandsophisticatedpowersolutions

Abstract

Onceverymuchasciencefictiondream,artificialintelligence(AI)israpidlybecomingpartofourdailylives.WhileAItakesmanyforms,thesystemsthatmimicthehumanbrain'slearningandproblemsolvingcapabilitycompriseincreasinglycapablecomputerbased“neural”networksconsistingofmanyparalleledprocessorsthatruncomplexlearningandexecutionsoftwarealgorithms.

Whilethealgorithmsarekeytothistechnology,thecomputerpowersystemdemandsarestretchingtheboundariesofexistingpowerdeliverytechnology.Inthiswhitepaper,InfineonwilllookatthepowerdemandsforAIsystems,aswellaspresentingsomeofthelatest,state-of-the-art,powertechnologiesthatareenablingtheadvancesinthisexcitingarena.

By

DannyClavette,Director,SystemsApplicationsInfineonTechnologiesAG

Artificialintelligence:poweringthedeep-learningmachinesoftomorrow

PAGE

10

04-2017

Tableofcontents

Abstract 1

Tableofcontents 2

Overviewofartificialintelligence 3

Challengesforpowerdesigners 4

Digitalvsanalogcontrol 5

RecentpowertechnologyadvancessupportingAI 6

Multi-rail&multiphasedigitalcontrollers 6

IR35411andTDA21470OptiMOS?powerstages 9

Summary 11

Overviewofartificialintelligence

Humansaresmart,achievingintelligencethroughyearsoflearninganddataaccumulationaswellasarguablygetting“wiser”withage.Computerscouldbeconsidered“smart”duetodataretentioncapabilitiesbutuntilrecentlylackedthecapabilitytoautonomouslylearnfromtheselargedatabasesinordertoexecutetasksormakedecisions.Whileahumanbrainconsumes20-30Wofpower,thelatestlearningsystemsareconsumingpoweratlevelsthatwouldsupportasmalltownastheylearntobecome'artificiallyintelligent'.Whilewecandebatewhethercomputingisgetting'smarter'thanhumans,itisimpossibletodebatethattherequirementsforpoweringthisnewgenerationofsupercomputerhavechangeddramatically.

Insomeways,theapproachtakentoAIdeeplearningisquitesimilartohumandevelopmentwherecomputerscontinuetolearnthroughexposure.Intheexamplebelow,aneuralnetworkisfedwiththousandsoftrainingimagesthatareprocessedviamultiplelayersinordertobuildexperienceandknowledge.

Asaresultofthiscomputerintensiveandpowerhungrylearningprocess,thenetworkiseventuallyabletodistinguishasquirrelfromachipmunkorafox.ThegoalistoachieveAIlearningintheshortestamountoftime,thusparallelcomputingpowerismaximizedtolinearlyimprovecomputationtimes.

Thehighpowerconsumptionoftoday'sAIisdrivingchangesinthecomputingarchitecturetoreplicateneuralnetworksthatmimicthehumanbraininanefforttoreducepowerneeds.TraditionalCentralProcessingUnits(CPUs)arearchitectedtobeveryflexibletosupportawidevarietyofgeneral-purposeprogramsandarenotoptimizedforveryspecificandrepetitivetaskssuchasAIlearning.

ManyofthenecessaryfunctionsforAIcanbeperformedbyGraphicsProcessingUnits(GPUs).TheseGPUsaredesignedtorepeatedlyperformcomplexmathematicalfunctionsmoreefficiently,canbeconvenientlyconnectedinparalleltofurtherincreasecomputingpowerandbeopportunisticallyappliedtolearningapplications.Withslightmodifications,theselatestGPUdevicesprocess3xto10xfasterwhileconsumingthesamepowerasaCPU.TheearlyAImarkethasbeendominatedbyNVIDIA;theirDX1GPUsupercomputercontainseightTeslaP100GPUs,eachcapableof21.2TeraFLOPs,andrequires3200Woftotalsystempower.MultipleDX1sconnectedinparallelarerequiredtoformaneffectiveneuralnetwork.

Honingthetechnologyevenfurther,TensorProcessingUnits(TPUs)areASICsthathavebeendevelopedspecificallyformachinelearning.BasedonGPUplatforms,reducedfloating-pointaccuraciesallowmorecomputecapabilityperclockcycle.Rasterizationandtexturemappingfeaturesarealsoremovedtofurtherimprovecomputationefficiency.GooglelaunchedthefirstTPUin2015andIntelisexpectedtolaunchLakeCrestthisyear,targetingDeepNeuralNetwork(DNN).

Tolearn,networksneedtobeabletosense.Local'edgedevices'includesensors,cameras,datacollectorsandlocalactuators.ConnectedtothecentralAIserversviahigh-speedwirelessconnections,theselowpowerdevicesaretheeyes,earsandhandsoftheneuralnetwork.Estimatespredictthattherewillbeover50billionedgedevicesconnectedtothenetworkby2020.

Itshouldcomeasnosurprisethat,despitethepowerchallenges,themarketforAIisgrowingrapidlyasdemonstratedbythe(approximately)40-foldgrowthatGoogleinthepasttwoyears.

Challengesforpowerdesigners

Thepowerlevelsrequiredforthisnewtechnologyaresimplystaggering.Inordertomatchtheprocessingpowerofahumanbrain,asystemwouldneedtoperformmorethan38thousandtrillionoperationspersecond(or38PetaFLOPSaccordingtoDharmendraModha,IBMFellowandChiefScientistattheAlmadenResearchCenter).Foraninterestingcomparison,aserverfarmusingNVIDIA’sDX1’s21.2TeraFLOPsper3200Wadvertisedperformancewouldrequireapproximately1800DX1sconsumingnearly6Megawatts(3200W*38e15/21.2e12).Thehumanbrainontheotherhand,requiresonly20Wofpower.

Thechallengefacingpowerdesignersismulti-faceted.Simplydeliveringtheselevelsofpowerischallengingenough.Efficiencyisabsolutelycritical,notonlyasenergycostsarerising,butalsoaseverywattofwasteenergydissipatedasheatincreasestheairconditioningrequirementsinthedatacenter,furtherincreasingoperationalcostsandcarbonfootprint.

Realestateisalsorisingincostand,asdatacenterscontainhundredsorthousandsofprocessingunits,sizeisimportant.Asmallreductioninthesizeofasingleunitisreplicatedmanytimesover,allowingmoreprocessingpowertobelocatedinthesamespaceaslargersolutions.Yet,thissmallersizerequirementrapidlyincreasespowerdensityandreducesthesurfaceareaavailablefordissipatingheat,makingthermalmanagementoneofthesignificantchallengesindesigningpowerforthisnewgenerationofAIsupercomputers.

Computingsystemsarecomplexloads;whilelearningtheyarerunningatfullpower.Astheactivitydrops,sodoesthepowerrequirement,buttheefficiencyisrequiredtoremainashighaspossiblethroughoutthepowerband.Withtoday'smulti-phasepowersolutionsthisentailsthedesignerbuildinginprovisionforcontrollingthenumberofphasesusedtoensurethatefficiencyisoptimizedatalltimes.

Digitalvsanalogcontrol

Clearly,amoresophisticatedapproachtopowerdesignisgoingtoberequiredtomeettheneedsofthisrapidlygrowingsector.Inordertoaddressthisneed,Infineonhasintroducedadvanceddigitalcontroltechniques,replacingthelegacyanalog-basedsolutions.

Digitalcontrolbringsmanybenefitswhendesigninghigh-endpowersolutions,notleastoverallsystemflexibilityandadaptability.Withdigitaltechnology,controllerscanbecustomizedwithouttheneedforexpensiveandtime-consumingsiliconspins.Thecustomizationextendstodefiningtheconfiguration,telemetryforgatheringsystemperformancedata,settingfaultmanagementandcalibratingthedevice.

Aspowersystemsbecomemoreintegratedintotheoverallsolution,communicationbetweenthepowersolutionandthemainCPU/GPU/TPUisanewrequirement.Infineon’smaturedigitalcontrollertechnologyfacilitatesmarket-leadingsolutionsandincludesaGUIthatenablesreal-timesystemdesign,configuration,validationandmonitoring.

DigitalsolutionssimplifybuildingthescalablepowersolutionsrequiredforAI.Yetwithalloftheincludedfunctionalityandprecisiondeliveryofpower,theyarenowpricecompetitivewiththeanalogsolutionstheyareultimatelyreplacing.

RecentpowertechnologyadvancessupportingAI

Infineonisoneoftheleadingdesignersandmanufacturersofadvancedpowercontrolandswitchingtechnologies.TheirproductofferingishighlyintegratedandprovidesallofthekeysiliconelementsrequiredtobuildhighlyadvancedpowersolutionsforAIapplications.

Infineon'scompleteportfolioincludeshugebreadthofproductsincludingdigitalcontrollers,integratedpowerstages,integratedpowermanagementICs,Point-of-Load(POL)convertersaswellasdiscretesolutionsincludingdriverICs,powerblocksanddiscreteMOSFETs.TherangeisbuiltuponInfineon'slonghistoryofinnovationandcomprisesmultiplemarket-leadingtechnologiessuchasOptiMOS?,DrMOS?andμDrMOS?.

Figure1 Infineonoffersfullflexibilityintermsofspace,performanceandcost

Multi-rail&multiphasedigitalcontrollers

CentraltoInfineon'sofferingforservers(aswellasworkstationsandhigh-enddesktops)isacompletecontrollerproductfamilyofmulti-rail/multiphasedigitalcontrollers.TheseadvancedcontrollersarecomplianttoIntel?,AMD?andsupportPMBUSwithAVS(AdaptiveVoltageScaling)forvoltageset-pointcontrolandsystemtelemetrywithupto50MHzmaxoperation.

Infineonsolutionsareprogrammabletoprovideone,twoorthreefullydigitallycontrolledvoltagerailswithupto10phases.InfineonalsooffersafamilyofdoublingICsandDriverstofurtherincreasephasecount.

Figure2 Infineon’srangeofadvancedcontrollersarecompliantwithIntel?andAMD?standardsandalsosupportPMBUSwithAVS(AdaptiveVoltageScaling)

Efficiencyacrossawideloadrangeissupportedthroughtheabilityofdesignerstoprogramautonomousphaseadditionorshedding.OtherprogrammablefeaturesincludePIDloopcompensation,loadlineslopeandoffsetaswellasdigitaltemperaturecompensation.

Externalloadlinesettingcomponentsareeliminatedbythedigitallyprogrammableloadline.ThecontrollerisdesignedtoworkwithRDS(ON)&DCRcurrentsensepowerstagesandprovidesaccurateinputandoutputcurrentreporting.

Digitalcontrolenablesproprietarynon-linearcontrolalgorithmsandprovidesexcellenttransientresponsewithreducedoutputcapacitance.Mostofourcontrollersalsosupportprogrammablecycle-by-cycleperphasecurrentlimitforsuperiordynamiccurrentlimiting.

ThesedevicesareeasilyconfigurableusingouroptimizedGraphicalUserInterface(GUI)toolswithfinalconfigurationsettingsthatcanbestoredinourdigitalcontroller’son-chipnon-volatilememory.

Aswouldbeexpectedofasophisticatedcontroller,significantfaultdetectionandprotectionisin-builtincludingIUVP,IOVP,CFP,OUVPandOOVP(InputUndervoltageProtection,InputOvervoltageProtection,CatastrophicFaultProtection,OutputUndervoltageProtectionandOutputOvervoltageProtection).Overcurrentprotection(OCP)isprovidedasaninstantaneousvalue,averagedfortotalcurrent,bychannelaswellaspulse-to-pulse.TherearemultipleOverTemperatureProtection(OTP)thresholds(internalandexternal)aswellasopen/shortvoltagesenselinedetectionandnegativecurrentlimitprotection.

InsomeofInfineon'slatestcontrollers,thecombinedstate-machineandintegratedmicrocontrollercorearchitectureallowformaximumflexibilityandtheinternalNon-VolatileMemory(NVM)storestheparametersofanycustomconfigurations.

Figure3 IR35215blockdiagram

IR35411andTDA21470OptiMOS?powerstages

TheIR35411powerstagecontainsalowquiescentcurrentsynchronousbuckgate-driverIC,high-sideandlow-sideMOSFETsandaSchottkydiodeinthesamepackagetofurtherimproveefficiency.ThepackageisoptimizedforPCBlayout,heattransfer,driver/MOSFETcontroltiming,andminimalswitchnoderingingwhenlayoutguidelinesarefollowed.ThepairedgatedriverandMOSFETcombinationenableshigherefficiencyatloweroutputvoltagesrequiredbycuttingedgeCPU,GPUandDDRmemorydesigns.

TheIR35411internalMOSFETcurrentsensealgorithmwithtemperaturecompensationachievessuperiorcurrentsenseaccuracyversusbest-in-classcontrollerbasedinductorDCRsensemethods.Protectionincludescycle-by-cycleOCPwithprogrammablethreshold,VCC/VDRVUVLOprotection,phasefaultdetection,ICtemperaturereportingandthermalshutdown.

Figure4 IR35411blockdiagram

TheIR35411featuresdeep-sleeppowersavingmode,whichgreatlyreducesthepowerconsumptionwhenthemultiphasesystementersPS3/PS4mode.

Operationofupto1.5MHzswitchingfrequencyenableshighperformancetransientresponse,allowingminiaturizationofoutputinductors,aswellasinputandoutputcapacitorswhilemaintainingindustry-leadingefficiency.

WhencombinedwithInfineon’sdigitalcontrollers,theIR35411incorporatestheBody-Braking?featurethroughPWMtri-statethatenablesreductionofoutputcapacitors.ThisquicklydisablesbothMOSFETsinordertoenhancetransientperformanceorprovideahighimpedanceoutput.TheIR35411isoptimizedforCPUcoreandmemorypowerdeliveryinserverapplications.

TheIR35411isanidealcompaniontotheIR35215multi-phasecontroller.

Figure5

belowshowshowtheIR35215combineswithfourIR35411stocreateaVRpowerstageina6+1configuration.

Figure5 VRusingIR35215controllerandIR35411powerstagein6+1configuration

Summary

WhileAIisstillearlyinitsdevelopment,itisalreadybeingrecognizedtobeanimportantandrapidlygrowingapplicationwithexpectedsubstantialimpactsonoursocieties.ThesepioneeringAIalgorithmsareenabledthroughseveralhighperformancecomputersystemsthatarechallengingdesignersonmanyfronts.Thetraditionaldatacenterdesignsarerapidlymigratingfromgeneral-purposeCPU-onlysolutionstowardscombinationsofCPUsandGPUsorTPUs,bringingnewandmorestringentdemandsondesignofserverpowersolutions.

Infineonoffersindustry’shighestefficiencypowerstagesthatutilizeInfineon’smarketleadingOptiMOS?MOSFETtechnology.ThroughcontinuedInfineonadvancementsinitspowersemiconductortechnology,ourdevicesarebecomingincreasinglyefficientresultingincontinuedpowerlossreductionswhileincreasingoursolutiondensities.

Infineondigitalcontrollersbringunprecedentedflexibilityandadaptabilityaswellasprecisecontrol,telemetryandprotectionfeatures.AsaleaderinthisAIpowerdeliverymarket,InfineonoffersabroadrangeofcontrollersandOptiMOS?powerstagesthatcansupportallknownAIhardwareplatformsandtheirdemandingcurrentlevels.Infineonenablesdesignerstocreatestate-of-the-artpowersolutionswithhighestefficiencyandpowerdensityfort