【高通Qualcomm】2025AI变革正在推动终端侧推理创新研究报告_第1页
【高通Qualcomm】2025AI变革正在推动终端侧推理创新研究报告_第2页
【高通Qualcomm】2025AI变革正在推动终端侧推理创新研究报告_第3页
【高通Qualcomm】2025AI变革正在推动终端侧推理创新研究报告_第4页
【高通Qualcomm】2025AI变革正在推动终端侧推理创新研究报告_第5页
已阅读5页,还剩14页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

AIdisruptionis

drivinginnovationinon-deviceinference

HowtheproliferationandevolutionofgenerativemodelswilltransformtheAIlandscapeandunlockvalue.

February2025

SnapdragonandQualcommbrandedproductsareproductsofQualcommTechnologies,Inc.and/oritssubsidiaries.

2

Contents

Executivesummary 3

QualityAImodelsarenowabundantanda<ordable 4

Innovationsboostmodelqualityandreducedevelopmenttimeandcost 4

Smallmodelsachievebigcapabilitiesattheedge 5

TheeraofAIinferenceinnovationishere 7

QualcommissettobealeaderintheAIinferenceera 8

Expandingacrossallkeyedgesegments 9

Mobile 9

PCs 10

Automotive 10

IndustrialIoT 11

Networking 11

Conclusion 11

3

Executivesummary

TheintroductionofDeepSeekR1,acutting-edgereasoningAImodel,hascausedripplesthroughoutthetechindustry.That’sbecauseitsperformanceisonparwithorbetterthanstate-of-the-artalternatives,disruptingtheconventionalwisdomaroundAIdevelopment.

Thispivotalmomentispartofabroadertrendthatunderscorestheinnovationincreatinghigh-qualitysmalllanguageandmultimodalreasoningmodels,andhowthey’repreparingAIforcommercialapplicationsandon-deviceinference.Thefactthatthesenewmodelscanrunondevicesacceleratesscaleandcreatesdemandforpowerfulchipsattheedge.

Drivingthisshiftarefourmajortrendsthatareleadingtoadramaticimprovementinthequality,performance,ande<iciencyofAImodelsthatcannowrunondevice:

•Today’sstate-of-the-artsmallerAImodelshavesuperiorperformance.NewtechniqueslikemodeldistillationandnovelAInetworkarchitecturessimplifythedevelopmentprocesswithoutsacrificingquality,allowingnewmodelsto

outperformlargeronesfromayearago,whichcouldonlyoperateonthecloud.

•Modelsizesaredecreasingrapidly.State-of-the-artquantizationandpruning

techniquesallowdeveloperstoreducethesizeofmodelswithnomaterialimpactinaccuracy.

•Developershavemoretoworkwith.Therapidproliferationofhigh-qualityAI

modelsmeansfeaturesliketextsummarization,codingassistantsandlive

translationarecommonindeviceslikesmartphones,makingAIreadyfor

commercialapplicationsatscaleacrosstheedge.

•AIisbecomingthenewuserinterface.PersonalizedmultimodalAIagentswillsimplifyinteractionsandproficientlycompletetasksacrossvariousapplications.

QualcommTechnologiesisstrategicallypositionedtoleadandcapitalizeonthetransitionfromAItrainingtolarge-scaleinference,aswellastheexpansionofAIcomputational

processingfromthecloudtotheedge.Thecompanyhasanextensivetrackrecordin

developingcustomcentralprocessingunits(CPUs),neuralprocessingunits(NPUs),

graphicsprocessingunits(GPUs),andlow-powersubsystems.Thecompany’s

collaborationwithmodelmakers,alongwithtools,frameworks,andSDKsfordeployingmodelsacrossvariousedgedevicesegments,enablesdeveloperstoacceleratethe

adoptionofAIagentsandapplicationsattheedge.

TherecentdisruptionandreassessmentofhowAImodelsaretrainedvalidatesthe

imminentAIlandscapeshifttowardslarge-scaleinference.Itwillcreateanewcycleof

innovationandupgradeofinferencecomputingattheedge.Whiletrainingwillcontinueinthecloud,inferencewillbenefitfromthescaleofdevicesrunningonQualcomm®

technologyandcreatedemandformoreAI-enabledprocessorsattheedge.

4

QualityAImodelsarenowabundantanda9ordable

Innovationsboostmodelqualityandreducedevelopmenttimeandcost

AIhasreachedthepointwherethedropinthecostoftrainingAImodels,combinedwithopen-sourcecollaboration,ismakingthedevelopmentofhigh-qualitymodelsaccessibletomorepeopleandorganizations.

Thisshiftisdrivenbyvarioustechnicaladvancements.Usageoflongercontextlength,

alongwithsimplificationofsomeofthetrainingsteps,savescomputationalcosts.Newernetworkarchitecturesrangingfrommixture-of-experts(MoE)tostate-spacemodels(SSM)arepushingtheboundaryofwhatcanbeaccomplishedwithreducedcomputational

overheadandpowerconsumption.

NewerAImodelsalsointegrateadvancedmethodssuchaschain-of-thoughtreasoningandself-verification,enablingthemtoperformwellacrossvariouschallengingdomainslikemathematics,coding,andscientificreasoning.

Distillationisakeytechniqueinthedevelopmentofcapablesmallmodels.Itallowslargemodelsto"teach"smallermodels,transferringknowledgewhilemaintainingaccuracy.Theuseofdistillationhasledtoasurgeinsmallerfoundationmodels—manyofthemfine-

tunedforspecializedtasks.

Thepowerofdistillationisexemplifiedinfigure1.ThispresentsaverageLiveBenchresultscomparingtheLlama3.370BmodelwithitsdistilledDeepSeekR1counterpart.Thechartshowshowdistillationsignificantlyenhancesperformanceinreasoning,coding,and

mathematicstasksforthesamenumberofparameters.

5

Figure1:LiveBenchAIaveragebenchmarkresultscomparingMetaLlama70Bmodelwithitsdistilled

counterpartbyDeepSeek.Source:LiveBench.ai,Feb.2025.

Smallmodelsachievebigcapabilitiesattheedge

Smallermodelsareapproachingthequalityoflargefrontiermodelsduetodistillationandothertechniquesdescribedabove.Figure2showsbenchmarksfortheDeepSeekR1

distilledmodelscomparedtoleading-edgealternatives.DeepSeek-distilledversions

basedonQwenandLlamamodelsshowareasofsignificantsuperiority,particularlyintheGPQAbenchmark–achievingsuperiororsimilarscorescomparedtostate-of-the-art

modelssuchasGPT-4o,Claude3.5Sonnet,andGPT-o1mini.GPQAisacriticalmetricbecauseitinvolvesdeep,multi-stepreasoningtosolvecomplexqueries,whichmanymodelsfindchallenging.

6

Figure2:Mathematicandcodingbenchmarks.Source:DeepSeek,Jan.2025.

ManypopularmodelfamiliesincludingDeepSeekR1,MetaLlama,IBMGranite,Mistral

Ministralfeaturesmallvariantswhichoverdeliverintermsofperformanceand

benchmarksforspecifictasks,regardlessoftheirsize.Thereductionoflarge,foundationalmodelsintosmaller,efficientversionsenablesfasterinference,smallermemoryfootprintandlowerspowerconsumption–allwhilemaintainingahighbaronperformance,allowingdeploymentofsuchmodelswithindeviceslikesmartphones,PCs,andautomobiles.

Furtheroptimizations,likequantization,compressionandpruninghelpreducemodel

sizes.Quantizationlowerspowerconsumptionandspeedsupoperationsbyreducing

precisionwithoutsignificantlysacrificingaccuracy,whilepruningeliminatesunnecessaryparameters.

Thesetechnicaldevelopmentshaveledtoaproliferationofhigh-qualitygenerativeAI

models.AccordingtodatacompiledbyEpochAI(Figure3),morethan75%oflarge-scaleAImodelspublishedin2024featurelessthan100billionparameters.

7

Figure3:Numberoflarge-scaleAImodelspublishedbyyear,categorizedbynumberofparameters.Source:

EpochAI,Jan.2025.

TheeraofAIinferenceinnovationishere

Theabundanceofhigh-quality,smallermodelsisbringingrenewedattentiontoinferenceworkloads–whichiswhereapplicationsandservicesmakeuseofthemodelstoprovidevaluetobusinessesandconsumers.

QualcommTechnologieshasworkedontheoptimizationofnumerousAImodelsto

supportthecommercializationofthenewgenerationofAI-orientedCopilot+PCs.

Similarly,thecompanyhascollaboratedwithOEMssuchasSamsungandXiaomiinthelaunchofflagshipsmartphonesequippedwithmanyAI-enabledfeatures.

TheproliferationofAIinferencingcapabilitiesacrossdeviceshasenabledthecreationofgenerativeAIapplicationsandassistants.Documentsummarization,AI-imagegenerationandediting,andreal-timelanguagetranslationarenowcommonfeatures.CameraappsleverageAIforcomputationalphotography,objectrecognitionandreal-timescene

optimization.

Nextupisthedevelopmentofmultimodalapplicationswhichcombinemultipletypesofdata—text,vision,audioandsensorinput—todeliverricher,morecontext-awareand

personalizedexperiences.TheQualcommAIEnginecombinesthecapabilitiesofcustom-builtNPUs,CPUsandGPUstooptimizesuchtaskson-device,enablingAIassistantsto

switchbetweencommunicationmodesandgeneratemultimodaloutputs.

AgenticAIispositionedattheheartofthenextgenerationofuserinterfaces.AIsystems

8

arecapableofdecision-makingandtaskmanagementbypredictinguserneedsand

proactivelyexecutingcomplexworkflowswithindevicesandapplications.QualcommTechnologies’emphasisonefficient,real-timeAIprocessingallowstheseagentsto

functioncontinuouslyandsecurelywithinthedevices,whilerelyinguponapersonal

knowledgegraphthataccuratelydefinestheuser’spreferencesandneeds,withoutanyclouddependency.Overtime,theseadvancementsarelayingthegroundworkforAItobecometheprimaryUI,withnaturallanguageandimage,videoandgesture-based

interactionssimplifyinghowpeopleengagewithtechnology.

Lookingahead,QualcommTechnologiesisalsopositionedfortheeraofembodiedAI,inwhichAIcapabilitiesareintegratedintorobotics.Byleveragingitsexpertiseininferenceoptimization,QualcommTechnologiesaimstopowerreal-timedecision-makingfor

robots,dronesandotherautonomousdevices,enablingpreciseinteractionsindynamic,real-worldenvironments.

WhilenumerousAImodelsaretrainedinthecloud,distilledsmallermodelsareavailableforoperationandrunondevicesoftenwithinweeksordays.Forexample,withinlessthanaweek,DeepSeekR1-distilledmodelswererunningon

PCs

and

smartphones

poweredbySnapdragon®platforms.

Deployinginferencewithindevicesaddressesimmediacythroughreducedlatency,

enhancesprivacy,reliesonlocaldatatoprovideadditionalcontextandenables

continuousfunctionalityofAIfeaturesandapplications.Italsoreducescostsforusersand/ordevelopersbyavoidingfeesassociatedwithcloudinferenceservices.AllofthiscreatesincentivesforsoftwareandserviceproviderstodeployAIinferenceattheedge.

QualcommissettobealeaderintheAIinferenceera

Asaleaderinon-deviceAI,QualcommTechnologiesisstrategicallypositionedtoadvancetheAIinferenceerawithitsindustry-leadinghardwareandsoftwaresolutionsforedge

devices.Thesesolutionsencompassbillionsofsmartphones,automobiles,XRheadsetsandglasses,PCs,industrialIoTdevices,andmore.

QualcommTechnologieshasalonghistoryofdevelopingcustomCPUs,NPUs,GPUsandlow-powersubsystems,which,whencombinedwithexpertiseinpackagingandthermal

design,formthefoundationofitsindustry-leadingsystem-on-chip(SoC)products.

TheseSoCsdeliverhigh-performance,energy-efficientAIinferencedirectlyon-device.Bytightlyintegratingthesecores,QualcommTechnologies’platformscanhandlecomplexAItaskswhilemaintainingbatterylifeandoverallpowerefficiency—criticalforedgeuse

cases.

TounlockthefullpotentialofAIonitsplatforms,QualcommTechnologieshasbuiltarobustAIsoftwarestackdesignedtoempowersoftwaredevelopers.TheQualcommAI

9

Stackincludeslibraries,SDKs,andoptimizationtoolsthatstreamlinemodeldeploymentandenhanceperformance.DeveloperscanleveragetheseresourcestoefficientlyadaptmodelsforQualcommplatforms,reducingtime-to-marketforAI-poweredapplications.QualcommTechnologies’developer-focusedapproachacceleratesinnovationby

simplifyingtheintegrationofcutting-edgeAIfeaturesintoconsumerandenterpriseproducts.

Lastly,thecompany’scollaborationwithAImodelmakersacrosstheglobeandits

provisionofservicesliketheQualcommAIHubarecentraltoitsstrategyforscalingAI

acrossindustries.OntheQualcommAIHub,inthreesimplesteps,adevelopercan1)pickamodelorbringtheirownmodelorcreateamodelbasedontheirdata;2)pickany

frameworkandruntime,writeandtesttheirAIappsonacloud-basedphysicaldevice

farm;and3)usetoolstodeploytheirappscommercially.TheQualcommAIHubsupportsmajorlargelanguageandmultimodalmodel(LLM,LMM)families,allowingdeveloperstodeploy,optimize,andmanageinferenceondevicespoweredbyQualcommplatforms.

Withfeatureslikepre-optimizedmodellibrariesandsupportforcustommodel

optimizationandintegration,QualcommTechnologiesenablesrapiddevelopmentcycleswhileenhancingcompatibilitywithdiverseAIecosystems.Thiscollaborativeapproach

strengthensQualcommTechnologies’positionasaleaderinenablingscalable,real-timeAIapplications.

Expandingacrossallkeyedgesegments

QualcommTechnologiesuseson-deviceAItosupportmanyindustries,unlocking

businessvalueandsupportingnewuserexperiences,allenabledbyenhanced

performance,efficiency,responsivenessandprivacybyprocessingAIlocallyondevices.

Mobile

Snapdragonmobileplatforms,suchasthelatestSnapdragon8elite,areadvancingthe

capabilitiesofon-deviceAIbyenablingseveralcutting-edgemultimodalgenerativemodelsandagenticAItooperatenativelyonsmartphones.AIhasenhancedsmartphonefeaturesacrossvariouscategoriessuchascommunicationimprovement,generativeimageeditingtools,personalization,andaccessibility.On-devicegenerativeAIisbeingutilizedto

developmoreintuitive,user-centricfeaturesandtoautomatetasksinmobiledevices.

ThistrendtowardsAI-drivenfunctionalitiesisevidentinthelatestflagshipsmartphonereleasesfrommajormanufacturersutilizingSnapdragonplatforms,includingSamsung,ASUS,Xiaomi,Oppo,Vivo,andHonor.

10

PCs

SnapdragonXSeriesplatformswereinstrumentalindefiningthenewcategoryofAIPCs,

withbest-in-classcustomNPUcoresthatwerebuiltfromground-upforhighperformance,energyefficientgenerativeAIinference.ThisNPUisturbo-chargingWindowsapps,addingnewfeatures,boostingperformance,andenhancingprivacyandbatterylife.Developers

canrungenerativeAIinferenceon-device,offeringcutting-edgeCopilot+PCfeatureswhichdebutedontheSnapdragonXSeries.

Popularthird-partyappslikeZoom,Affinity,DjayPro,CapCut,MoisesLive,and

BlackmagicDesign’sDaVinciResolvetakeadvantageoftheNPUtoofferspecificAI-poweredcapabilitiesonSnapdragonXSeriesplatforms.

Automotive

Snapdragon®DigitalChassis™solutionuseson-deviceAIinitscontext-awareintelligentcockpitsystemdesignedtoenhancevehiclesafetyanddriverexperience.Thissystem

leveragesadvancedcameras,biometricandenvironmentalsensors,andstate-of-the-artmultimodalAInetworkstoprovidereal-timefeedbackandfunctionalitytailoredtothe

driver'sstateandenvironmentalconditions.

Forautomateddrivingandassistancesystems,QualcommTechnologieshasdevelopedanend-to-endarchitecturewhichuseslargetrainingdatasets,fastre-trainingusingreal-worldandAI-augmenteddata,over-the-airupdates,andastate-of-the-artstackincludingmultimodalAImodelsandcausalreasoninginthevehicletohandlemodernautomated

drivingandassistancecomplexities.

Example:LLMAgentlistenstheconversationsinthecabin,onepassengermentionscoffee,

afterafewminsPOIshowsacoffeehouse,LLMAgentproposesastopforcoffee)

Perception-to-IVI

LLMAgent

(AIAssistant)

EnhancedARHUD

Perception

In-VehicleSensors

IntuitiveHMI

Driving

MultimodalLLM

DeepPoints,POI

PlanningProposals&DriverStatus

HumanDriver

Effectivesceneunderstandingandcognition

Decision

ADAS

ADSensors

Transformer

Tokenized

Environment

EnvironmentTokenization

Improvedspatialreasoningandreal-timeplanningcapabilities

Perception-to-ADAS

11

Figure4:Simplifiedin-vehicleAIsystemarchitecturetosupportintelligentcockpitandautonomousand

advanceddrivingassistance.Source:QualcommTechnologies,Jan.2025,

IndustrialIoT

ForindustrialIoTandenterpriseapplications,QualcommTechnologiesrecently

introduceditstheQualcomm®AIOn-PremApplianceSolution,anon-premisesdesktoporwall-mountedhardwaresolution,andQualcomm®AIInferenceSuite,asetofsoftwareandservicesforAIinferencingspanningfromnear-edgetocloud.

ThisedgeAIapproachallowssensitivecustomerdata,fine-tunedmodels,andinference

loadstoremainonpremises,enhancingprivacy,control,energyefficiency,andlow

latency.That’scriticalforAI-enabledbusinessapplicationssuchasintelligentmulti-

lingualsearch,customAIassistantsandagents,codegeneration,andcomputervisionforsecurity,safetyandsitemonitoring.

Networking

QualcommTechnologieshasintroducedanAI-enabledWi-Finetworkingplatform–the

Qualcomm®NetworkingProA7Elite.ThesolutionintegratesWi-Fi7andedgeAItoallow

accesspointsandrouterstorungenerativeAIinferenceonbehalfofconnecteddevicesinthenetwork.Itsupportsinnovativeapplicationsinareaslikesecurity,energymanagement,virtualassistants,andhealthmonitoringbyprocessingdataonthegatewayforenhancedprivacyandrea

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论