mckinsey -人工智能時代的軟件定義硬件 Software-defined hardware in the age of AI

January2025

Mcsey

&company

Automotive&AssemblyPractice

Software-defined

hardwareintheageofAI

Software-definedhardwareisreshapingmultipleindustriesasadvancesinAIenablenewcapabilitiesanddramaticallyreducedevelopmentcosts.

byAliRizvi,AniKelkar,andPhilippKampshoffwithSarthakVaish

Overthepasttwodecades,theshiftfromfixed-functionhardwaretosoftware-defined

hardwarehasrevolutionizedindustriesrangingfromnetworkingtomobilecommunications.With

software-definedhardware,developerscanimproveproductsandservicesbycontinuallyupdatingsoftwareratherthanundertakingmorecostlyandtime-consuminghardware

upgrades.Devicesthatwereoncerigidandtaskspecificarenowbecomingprogrammableandflexible,allowingthemtohandlenewtasksanddemands.

Despitethefunctionalbenefitsofsoftware-definedhardware,itsusewastraditionallylimitedtoindustriesthatofferedsubstantialunitvolumes,suchassmartphones,toamortizethefixed

developmentcosts.Inthesesectors,physicallyreplacingorupgradingdevicesismuchmore

expensivethanmakingsoftwareupdates.(Seesidebar“Transformingindustriesthrough

software-definedhardware”formoreinformationonearlyusecases.)Industriesthathadlowerhardwarevolumesdidnotfeelthesameurgency,becausesoftwaredevelopmentcostswereinlinewithorexceededthosefordevicereplacement.Today,however,AIischangingthecost-

benefitbalancebyautomatingmanyroutinesoftwaredevelopmenttasks,therebyreducingthetimeandlaborrequiredandexpandingthefunctionalcapabilitiesofthesoftware.

Beyondimprovingefficiency,software-definedhardwaremayhelpcompanieswinnew

customersandincreasebrandloyaltybyimprovingdeviceperformanceandenablinggreater

Transformingindustriesthroughsoftware-definedhardware

Software-definedhardwarefirstdemonstrateditsabilitytotransformentireindustriesaround

2009,whendevelopersintroducedsoftware-definednetworking(SDN).Bydecouplingnetwork

controlfromthephysicalinfrastructure,SDNenableddynamictrafficmanagementandprovided

moreflexibilityinnetworkscalinganddeployment.Manycompanies,includingCisco,VMware,andJuniperNetworks,havealreadyshiftedtoSDNintheirdatacenters.

Inthe2000s,thesmartphonerevolutionbroughtthebenefitsofsoftware-definedhardwaredirectlytoconsumers.Unlikeearliermobilephoneswithfixedcapabilities,smartphoneplatformscouldbe

continuouslyredefinedbyappdevelopers,blurringthelinebetweenhardwareandsoftware.By2012,smartphoneappscouldallowuserstocustomizetheirdevicesforGPSnavigation,photoediting,andothertasks.

Software-definedhardwareintheageofAI2

personalization.Acar’sinfotainmentsystem,forinstance,couldprovidecustomized

entertainmentandstreamingoptionsbasedonthedriver’spreviouschoices.Ifthesoftware-definedhardwareincorporatesAIormachinelearning(ML)algorithms,itmaytakeproduct

performancetoevengreaterheightsandlearnfrominteractionswithcustomers.

AsAIenablesfurtherproductdevelopmentcostreductions,evenmoreindustries,includingaerospace,medicaldevices,andconsumerdevices,couldacceleratescalingofsoftware-

definedhardware.First,however,theymustupdatetheirorganizationalstructuresand

operationstomaximizeAI’sbenefits.Here’salookatsomechangesthatmayhelp,withafocusonexamplesfromtheautomotiveindustry.

BenefitsofAIinsoftware-definedhardware

Formanycompanies,theadvancesinAIhavecomeatanauspicioustime.Softwarecomplexityhasbeenincreasingsteadilyandcouldriseevenfurther.Considertheautomotivesector:since2021,thecomplexityoftheaveragevehiclesoftwareplatformandthetotaleffortrequiredto

createithavebothincreasedbyabout40percentannually(exhibit).Overthesameperiod,however,softwaredevelopmentproductivityhasincreasedonlybyabout6percentperyear.

UsingAItooptimizesoftware-definedhardwaredevelopment

AdvancesinAIareexpandingthetoolboxofproductdevelopersandimprovingthedevelopment

ofsoftware-definedhardwarebyautomating,optimizing,andenhancingvariousstagesofdesign,development,andtesting.

AI-assisteddesign.SeveralgenerativeAI(genAI)toolscanhelpengineersdevelophardware

descriptionsandlayouts,reducingmanualeffort.Forinstance,onetooloptimizeshardware

architecturedesignbasedonspecifiedconstraintsandobjectives.Othernew

generativedesign

systemscanexploreamuchlargeruniverseofpossiblesolutionsthanthepreviousgenerationoftools.Bycomparingtheresultsofthousandsofsimulations,theycanhelpidentifyadesignthatdeliversthemostfavorablecombinationofattributes.

Software–hardwarecodevelopment.UsingAIagentsduringproductdevelopmentcanhelp

bridgegapsbetweensoftwareandhardwaredesignbyensuringconsistentrequirementsacrossiterations.AIagentscanalsoensurethatthehardwareiseasytoprogramandthesoftwareis

optimizedforhardwareperformance(forinstance,AIcanadjustsoftwareroutinestoimproveutilizationofGPUs,specializedaccelerators,orotherresources).

Hardwareoptimization.AIalgorithmscanoptimizetheallocationofresources,suchasmemory,logicblocks,andprocessingunits,forsoftware-definedhardwaretomeetobjectivesandenableadaptationsifandwhenproductrequirementsevolve.AIcanalsoidentifyissuesandflawsearlyinthedesignprocessthroughtheuseofsimulationsandimproveproductreliabilityand

Software-definedhardwareintheageofAI3

Exhibit

vehiclesoftwareplatformcomplexityandtheeffortrequiredtocompletedevelopmentprojectsareincreasing,whileproductivityremainsflat.

Automotivesoftwarecomplexity,effort,andproductivity,index(2021=100)

DesigncomplexityprojecteffortDevelopmentproductivity

800700

500400300200

CAGR:42%o

CAGR:36%

CAGR:6%

800700

500400300200

202120242027202120242027202120242027

source:Basedonanalysisof280software-intensiveautomotiveplatformprojectsintheNumetricssoftwareindustrydatabase

Mckinsey&company

security.Hardwareoptimizationcancontinuebeyondtheinitialdesignstagebecauseedgeprocessingallowsengineerstomanageresourcesdynamicallytomeetobjectives.

Acceleratedtesting.Deep-learningsurrogatesnowallowengineerstoreplacemanyphysicaltestswithfastervirtualassessmentsthatcostless.

WithAIhandlingthesetasks,morecompanieswillbeabletocreatesoftware-definedhardwareforthefirsttimeoraccelerateinnovation.London-basedtechnologycompanyWayve,for

example,isincorporatingitsLINGO-1genAImodelintoitsself-driving-carsoftwaretoanalyzedriverinputsanddatafromvehiclesensors,includingimagescapturedonvideo.1Using

LINGO-1capabilities,thevehicle’ssoftwarewillthengenerateanswerstocommonquestions,

suchas“Whydidyoustopatthispoint?”Similarly,WaymousesCarcraft,asoftwareprogram

thatmodelshowvirtualcarsnavigatedifferenton-roadscenariosinactualcities,toimprove

autonomousdriving.2Thevirtualvehiclestravelabouteightmillionmilesperday,muchofthisincomplicatedtrafficsituations,suchasrotaries.

1“UnveilingLINGO-1’sshowandtellcapabilitywithreferentialsegmentation,”Wayve,November15,2023.

2AlexisC.Madrigal,“InsideWaymo’ssecretworldfortrainingself-drivingcars,”Atlantic,August23,2017.

Software-definedhardwareintheageofAI4

UsingAItocreatebetterendproductsandservices

Thebenefitsofsoftware-definedhardware,enhancedbyAI,translateintobetterproductsandservicesforendcustomers.Asanexample,comparethecustomerexperienceatelectricvehicle

(EV)start-upswiththatoftraditionalOEMs.NewEVOEMs,suchasTesla,whichwerenot

constrainedbylegacyhardware,haveusedsoftware-definedhardwareanddeliveredover-the-air(OTA)updatestoimprovevehicleperformance,efficiency,andfeaturesbeginningin2018.

Thatconvenienceisnotpossiblewitholder,hardware-centriccardesigns.OtherOEMsarealsoinvestinginsoftware-definedhardwareorformingjointventurestocreateit.

Hereareafewotherwaysthatsoftware-definedhardwareandAIcanimproveendproductsandservices.Notethattheseenhancementsrelatetoproductsthatcompaniessellto

consumersorotherbusinesses.Companiesmayalsousesoftware-definedhardwaretoimprovetheirownoperations(seesidebar“Makingawarehousemoreefficient”).

Optimalhardwareperformanceandlifetime.AIcananalyzedatafromdevices,suchascarsandmedicalequipment,tooptimizesoftwareconfigurationsforbothefficiencyandperformance.

Forinstance,onenewEVOEMusedAItoreducethenumberofelectroniccontrolunitsinitsvehiclesfromasmanyas80insomelegacysystemstounderten.

Personalizeduserexperiences.GenAIcanmakesoftware-definedsystemsmoreintelligent.In

Makingawarehousemoreefficient

Althoughcompaniesmaychoosetofocusondevelopingsoftware-definedhardwareforconsumerorbusinessclients,theycanalsousethemtoimprovetheirinternaloperations.WiththeadditionofAIcapabilities,software-definedhardwarecandeliverfarmorethanitcouldevenafewyearsago.

Considerawarehousesettinginwhichhumansinteractregularlywithvarioustypesofrobots.Duringanidealinteractionforautomatedpicking,arobotindicatestheitemtobepickedanditslocation.

Thewarehouseassociatethenpicksuptheitem,scansit,andplacesitinatote.Butwhenthereisaproblem—forinstance,anitemisdamagedormissing—theworkerhastoflagtherequestasanerrorthatneedstoberesolvedbyacommandcenterteam,resultinginlostproductivity.Human–machineinterfacesthatincorporategenerativeAIwouldalreadyhaveaccesstotheinformationneededfor

problem-solving,allowingtheworkertoaddresstheissuequickly.

Software-definedhardwareintheageofAI5

smarthomes,forinstance,genAIsystemscouldlearnahousehold’shabitsandautomaticallyadjusteverythingfromlightingtosecurity.Ifthesystemsreliedontraditionalhardware,ratherthansoftware-definedhardware,suchadjustmentswouldbedifficultandexpensivetomake.

Manymedicaldevicesarebecomingincreasinglysoftwaredriven,withcompaniessuchas

MedtroniccreatingAI-enhancedsystemsthatadaptbasedonpatientdata,suchasvitalsigns.

Intheconsumersector,devicessuchasAmazonEchoalreadyrelyonsoftwaretointroducenewfeatures,improvetheirintelligence,andenhancetheuserexperience.AsmorecompaniesbeginincorporatingAIintotheirproducts,theycanusesoftware-designedhardwaretoprovidemorepersonalizedservice.Forinstance,severalnewentrantshavelaunchedgen-AI-enableddigital

personalassistants.

Betterhuman–machineinterfaces.Human–machineinterfaces(HMIs)havelongallowed

peopleto

interactwithdifferentdevices

,rangingfromindustrialmachinestopersonaldevicestoservicerobots,butthecommunicationchannelscanbefrustratingorimperfect.UsingAI,somecompaniesareattemptingtoimprovethem.Withintheautomotivesector,forinstance,the

locationtechnologycompanyTomTomhasdevelopedanAIin-vehicleassistantthatappliesrecentadvancesinAItoengageincomplex,multifacetedconversations.Thevoicesystemproactivelygivesdriversusefulinformation,ratherthanwaitingforthemtoaskquestions.

BMW’sPanoramicVisionisanewheads-updisplaythatprojectsinformation,suchasvehiclespeed,ontowindshields.ItusesAItotrackthedriver’seyemovementstoensurethat

informationalwaysappearsinaclear,easy-to-seelocationthatwillnotinterferewiththeviewoftheroad.Themostimportantinformationappearsinadarkenedsectionatthebaseofthe

windshield,whilelessimportantdataareprojectedslightlyhigherontoaclearsection.The

technologycanautomaticallydeterminewhencertaininfoshouldbedisplayedinapriorityspot.Forinstance,itprominentlydisplaysnavigationalinformationtohelpwithparkingwhenadriverislookingfororenteringaspot.

Applicationoffoundationmodels.Foundationmodelsaretrainedonvastdatasetsand

leverageadvanced-machine-learningtechniquestocreatebetterhardware.Theyhavealreadybeenusedtooptimizeperformanceandfunctionalityinenterpriseandconsumerproducts.

EmbodiedAI,whichletsrobotsandotherintelligentdevicesinteractwiththeirenvironmentandlearnfromexperience,couldbethenextimportantapplication.Somefoundation-model-basedroboticsstart-upsarehopingtocreatemoregeneralizedhardwarewithbroaderfunctionality

thatimprovesovertime.(Anewapproachtomodelcreationmayhelpexpeditedevelopment,asdescribedinthesidebar“Whataremodularfoundationmodels?”)

Anewwayofworking

HarnessingAI’spowermayrequirecompany-wideoperationalchangesorindustry-wide

Software-definedhardwareintheageofAI6

Whataremodularfoundationmodels?

Withmodularfoundationmodels,AIcapabilitiesaredividedintosmall,distinctunits.Developerscanconfigurethemodulestoworktogetherindifferentways,muchlikehowdifferentpartsofthehumanbraincommunicate,andadjustthemwhenneeded.Themodulardesignmaybeparticularly

helpfulforlargelanguagemodels,whichcanunderstandandanalyzetext.Advantagesofconfigurablemodelsincludethefollowing:

—Efficiency.Byactivatingonlythenecessarymodulesforagiventask,modularfoundationmodelssignificantlyreducecomputationaloverheadandenergyconsumption.

—Adaptability.Modulescanbeadded,removed,ormodifiedtoadapttonewtasksandevolvingrequirements,makingthemodelshighlyflexible.

—Scalability.Modulardesignenablesincrementalimprovementsinperformance,allowingforgradualscalingandeliminatingtheneedforcompleteretraining.

—Privacy.Bypartitioningamodelintosmallmodulesthatperformspecializedtasks,sensitiveinformationcanbeisolated,enhancingprivacyandsecurity.

initiativesrelatedtocollaboration,marketentry,decoupledhardwareandsoftwaredevelopment,anddatasecurityandprivacy.

Innovativecollaborations

AsmorecompaniesbeginusingAItocreatesoftware-definedhardware,theyshouldensure

thathardwareengineersandsoftwaredeveloperscollaboratefromtheprogram’soutset.Theyalsoshouldensure,veryearlyon,thatthehardwaredesignisflexibleenoughtoallowsoftware-designedhardwaretomakeupdates.

Insomecases,companiesmayformnontraditionalpartnerships,suchaswithtechstart-ups,tocreatesoftware-definedhardware.Theycanfacilitatesuchcollaborationbymandatingtheuseofstandardprotocolsthatallowdynamicmanagementofnetworkdevicesusedinsoftware-

definedhardware,includingtheOpenFlowandtheNetworkConfiguration(NETCONF)

protocols.Whilestandardprotocolscanbeusefulinanycollaboration,theyareparticularly

importantwhentwopartnershaveneverworkedtogetheranddonotunderstandhowtheothercompanyoperates.

Someprotocolsareindustryspecific,includingtheAutomotiveOpenSystemArchitecture

Software-definedhardwareintheageofAI7

(AUTOSAR),whichprovidesarobustframeworkfordevelopingsoftware-definedautomotivehardware.AUTOSARofferstwocomplementarystandards:ClassicAUTOSAR,whichiswellsuitedfortraditionalembeddedsystemswithstrictreal-timerequirements,andAdaptive

AUTOSAR,whichisdesignedformorecomplex,dynamic,andsoftware-intensivefunctions.

AdaptiveAUTOSARenablesthecreationofscalable,modular,andsecureapplications,someofwhichmayinvolveAIfunctionalities.Forinstance,AdaptiveAUTOSARiswellsuitedtomanaging

advanced-driver-assistancesystems(ADAS),whichmayincludeAI-basedperceptionand

decision-making.AdaptiveAUTOSARalsosupportsthedeploymentofapplicationsthatinteractwithradar,ultrasonic,andothersensorsessentialforADAS,althoughspecializedAIframeworksmightstillberequiredtohandleothertasks,suchasdataprocessing.

Otherprotocolsthatmayacceleratethedevelopmentofsoftware-definedhardwareincludetheSafety-OrientedArchitectureforFunctionalSafetyinEdgeComputingprotocol,sponsoredbytheScalableOpenArchitectureforEmbeddedEdge(SOAFEE)initiative,whichfocuseson

creatingsafetystandardsforedgedevices.

Byfollowingthesecommonframeworks,developerscanbuildflexible,scalable,andefficientsystemsthatadapttoevolvingtechnologicaldemands.

Astheuniverseofpartnersexpandsanddiversifies,companiescanencouragemoreAI-relatedinnovationbyloweringbarrierstomarketentry.Forinstance,theycouldeliminatelicensingfeesforstart-upsthatwanttodevelophardwarebasedontheirtechnology.Overall,companies

shouldfocusonenablingpotentialpartners,ratherthancontrollingtheiractivities.

Decoupledhardwareandsoftwaredevelopmentcycles

Tocreatesoftware-definedhardware,companiesshoulddecouplesoftwareandhardware

developmenttoenablegreaterflexibilityandinnovation.Thisdoesnotmeanthateachelementevolveswithnoinputfromtheothergroup,sincecollaborationismoreimportantthaneverwithsoftware-definedhardware.Itdoesmean,however,thathardwareandsoftwareeachfollow

theirowntimelinesandupdatecycles.Thepacesetbythetwogroupswillnotnecessarilybeinsync.

Decouplingallowshardwareengineerstodesignandoptimizephysicalcomponents

independently,whilesoftwaredeveloperscancreateandupdateapplicationswithoutbeingconstrainedbyhardwarelimitations.Byutilizingabstractionlayersandstandardinterfaces,teamscaniterateonsoftwarefeaturesandfunctionalitiesrapidly,fosteringamoreagile

developmentprocess.Decouplingalsofacilitateseasierintegrationofnewtechnologiesand

upgrades,ashardwarecanevolvetomeetchangingdemandswithoutrequiringacompleteoverhaulofthesoftwarestack.Ultimately,thisapproachenhancescollaborationacross

disciplinesandacceleratestimetomarketfornewproductsandsolutions.

LegacyautomotiveOEMsareamongthosedecouplinghardwareandsoftwaredevelopment,

Software-definedhardwareintheageofAI8

Findmorecontentlikethisonthe

MckinseyInsightsAPP

scan·Download·personalize

partlyb