2023人工智能白皮書：為AI奠基未來（英文版）

PREPARINGTHE

FOUNDATIONFORTHEAIOFTOMORROW

LianJyeSu,ResearchDirectorMalikSaadi,VicePresident,StrategicTechnologies

CONTENTS

EVERYAIINVESTMENTHASTOBEA

FORWARD-LOOKINGBUSINESSDECISION

The2020sareshapingupasthedecadeofArtificialIntelligence(AI).Withtherightinvestments,thetechnologyispoisedtobewidelydeployedinvariousInformationTechnology(IT)andOperationalTechnology(OT)usecases.PwC’sGlobalAIStudy:ExploitingtheAIRevolutionestimatesthatAIwillcontributeUS$15.7trilliontotheglobaleconomyby2030.

Notsurprisingly,multipleAIadoptionstudiesconductedbymajorAIplayers

revealthatmanybusinesseshaveincreasedtheirAIbudgetquitesignificantly

comparedtopreviousyears.However,mostinvestmentdecisionsaredrivenbythebroadpromiseofAIwithoutfocusingonspecificimplementationhurdlesinlargeorganizationswithlegacyfootprints.Asaresult,manybusinessesriskcreatinga

siloed,looselyintegrated,andproprietarysystem.

ThiswhitepaperaimstounpackthedifferentfacetsofAIandtheirrespective

computationalrequirements,showingthatAIinvestmentsmustbebasedonlong-termbusinessoutcomesandvalues.

EVERYAIINVESTMENTHASTOBEA

FORWARD-LOOKINGBUSINESSDECISION 1

MULTIMODALITYISTHEFUTUREOFAI 2

DIVERSITYINAIIMPLEMENTATION

ENVIRONMENTS 3

PRIVACYANDSECURITY-ENHANCEDAI 4

PREPARINGFORTHEAIOFTOMORROW 4

KEYPRINCIPLESOFAIINFRASTRUCTURE

INVESTMENT 7

CHARACTERISTICSOFAFUTURE-PROOFED

AIINFRASTRUCTURE 8

COMPREHENSIVEANDHETEROGENOUS

INFRASTRUCTURE 11

EDGE-TO-CLOUDVISION 13

OPENNESS 13

SECURITYINFUSEDATEVERYLAYER 14

BACKWARDCOMPATIBILITY 14

KEYTAKEAWAYSANDRECOMMENDATIONS

FORENDUSERS 15

TAKEAWAY1:DEVELOPACLEARINTERNALAI

ROADMAPBASEDONBUSINESSOUTCOMES 15

TAKEAWAY2:GETORGANIZATIONALBUY-IN 16

1

CommissionedbyIntelCorporation.

PREPARINGTHEFOUNDATIONFORTHEAIOFTOMORROW

2

PREPARINGTHEFOUNDATIONFORTHEAIOFTOMORROW

TAKEAWAY3:FOCUSONSOLUTIONPROVIDERS

ANDCHIPSETSUPPLIERSTHATEMBRACE

OPENNESS,FREEDOMOFCHOICE,TRUST,

ANDSECURITY 16

TAKEAWAY4:LEVERAGESUPPORT

FROMECOSYSTEMPARTNERS 16

CASESTUDY:INTEL 17

MULTIMODALITYISTHEFUTUREOFAI

Beforedivingintotheinfrastructurerequirements,itisworthwhilehavingan

overviewoftoday’sAI.Asimplementedtoday,AIgenerallyfocusesonfourmajorapplications,asdescribedinFigure1.

Figure1:MajorClassesofAIofToday

(Source:ABIResearch)

LANGUAGE

VISION

TABULAR

GRAPH

MULTIMODALAI

AllNaturalLanguageProcessing(NLP)tasks,suchas

Acombinationoftwoorthreedatasourcefor

recommendersystem,conversationalAI,etc.

machinetranslation,sentimentanalysis,speech

enhancement,speechsynthesis,andwakeworddetection.

Facialrecognitionforaccesscontrol,OpticalCharacter

Recognition(OCR),presencedetection,onjectandlandscapedetectionincomputationalphotography,andvideorecognitioninpublicsafetyandtrafficoptimization.

Informationfromvarioussensors,includingpressure,

temperature,vibration,rotation,acceleration,force,andacoustics,tomonitormachines’status,predictequipmentperformance,andtrackimportantassets.

Datathatareconnectedtooneanotherthroughspecific

relationship,suchassocialnetworkgraph,oraresequentialanddemonstratiespecificpatternsinaprocess,suchas

e-commerceconsumptionoronlinesearchdata.

AIapplications,suchasRoboticsProcessAutomation(RPA)andOpticalCharacterRecognition

(OCR),arealreadyautomatingmundaneandrepetitiveworkflows,helpinghumanemployees

performbetterattheircurrenttasksandassistingbusinesseswithstayingcompliantwithlegalrequirements.However,theAImodelsintheseapplicationsarefocusedonahighlystructuredworkloadusingoneofthefourdatatypesmentionedabove:language,vision,tabular,orgraph.Therefore,asignificantportionoftomorrow’sAImodelswillbemoreversatile,leveragingmost,ifnotall,ofthefourdatatypes.

TheseAImodelsaredesignedforavarietyoftasks.Theycanimprovelearning,decision-making,andexperiencesbyrelyingontrainingfromvariousdatasources,recordsandarchives,and

personalinformationwiththeindividual’sexplicitconsent.However,beforethisvisionbecomesareality,theindustryneedstodevelopmorerobustmultimodallearningmodelstoprocess

variousdatainrealtime.Thesemodelswillalsoneedtobesupportedbyhigh-performance

heterogeneouscomputingarchitecturetoenableawiderangeoftasksandfunctions,fromdatagatheringandstructuring,whichconsumesasignificantshareofprocessingresourcestoreal-timeinferenceandothermissioncriticaltasksthatrequireultra-lowlatencyandhighaccuracy.

DIVERSITYINAIIMPLEMENTATIONENVIRONMENTS

AItechnologysuppliersanddevelopersarealsotakingadvantageofnewAItechniquesand

edgecomputingtechnologytodeployAIacrossalldevicesfromthecloudtotheedge.Ononehand,CloudServiceProviders(CSPs)areleveragingscalableandhyperscaledatacentersto

developanddeployhigh-performancevision,language,andgraphmodelsthatareconstantly

increasinginsize.Ontheotherhand,businessesarelookingforAImodelsembeddedindevicesandgatewaystoimprovelatency,protectprivacy,andreducerelianceoncloudinfrastructure.

TinyMachineLearning(TinyML)pushestheboundaryfurtherbyintroducingultra-lowpowerAIinferenceinsensorsandbattery-powereddevices.SoftwarecapabilitieslikeNeuralArchitecturalSearch(NAS)andnewmodelcompressiontechnologieslikeknowledgedistillation,pruning,andquantizationwillenableAIdevelopersandimplementerstocreatethemostoptimizedmodelfortheirtargetenvironment.

AlltheseadvancementsmeanAImodelswillbepresentineverynodeofthecomputing

continuum,rangingfromhyperscaledatacenterstoregionaldatacenters,on-premisesservers,edgecomputinggateways,devices,andsensors.AlltheselocationsenablebusinessestodeployAIatthemostoptimallocationintermsofcomputingpower,latency,connectivity,andregulation.

Inaddition,thearrivalofnext-generationtelecommunicationtechnologies,suchas5Gand

Wi-Fi6,alsoallowsthetransferofalargeamountofdatafortrainingandinference.Whilesuchinfrastructureisnotcurrentlyavailableineverylocation,thisrealitywillchangeascloudAIgiants,telecommunicationserviceproviders,industrialcompanies,andedgecomputingcompanies

continuetobuildtherelevantinfrastructureinthenextfewyears.Asaresult,AIdeveloperswillwidelyintroduceAImodelsbasedonnewtechniques,heterogeneoushardware,andlowlatencyconnectivity.

Figure2:Edge-to-CloudComputingContinuum

(Source:ABIResearch)

Hyperscaledatacenter

Massive-scale

datacenter

builttosupport

alltypesof

workloadsina

centralized

location.

Regionaldatacenter

Regionaldata

center,including

hyperscaler

backbone

andservice

extensionsinto

theenterprise.

Country/metrodatacenter

Country-or

city-leveldata

center,including

hyperscaler

backbone

andservice

extensionsinto

theenterprise.

RadioAccessNetwork(RAN)

Equipmentthat

enablesnetwork

operatorsto

connectcustomer

assetstothe

network.Alsohosts

multi-accessedge

computeserver.

On-premisesserver

Serversthatare

locatedon

customers’

premises

tosupport

alloperational

workloads.

Gateway

Ahubthat

connectsto

multipledevices

andperformsdevicemanagementandconfigurations.

Deviceandsensor

Assets,machinery,toolsandequipment

thatfeature

internet

connectivityand

arelinkedtoa

publicand/or

3

privatedatacenter.

CLOUDCENTRICEDGECENTRIC

PREPARINGTHEFOUNDATIONFORTHEAIOFTOMORROW

PRIVACYANDSECURITY-ENHANCEDAI

AsidefromclassicalAIandsomeMachineLearning(ML)models,mostoftoday’sDeepLearning(DL)modelsareblackboxesthatlacktransparency.AIdevelopersdonothavecomplete

knowledgeofalltheindividualneurons,layers,andparametersinaDLmodelthatworktogethertoproducefinaloutput.Theroleofallthesecomponentsandtheirinfluenceovereachother

remainslargelyunexplained.

Movingforward,AIdeveloperswillmakeAImodelsmoretransparent.Inmostcases,adata

andAIdevelopmentplatformwillbedesignedtoexplaintousersthelimitationsoftrainingandtestingdata,thelogicbehindallAItrainingandinferenceprocesses,andpotentialbias,drift,andothergaps.Insomecases,AImodelsmaybedesignedtoexplainthemselvestotheendusers.ThetransparencyandexplainabilitywillenableAImodelstobeusedinhigh-riskenvironments,astheycanwithstandscrutinyandevaluation.

Furthermore,AIdeveloperscanenhancesecurityinanAIsystembylimitingdatatransfertothecloudanddeployingtheAImodelattheedge.Bykeepingandprocessingrawdataattheedge,endusersdonotneedtoworryabouttheirdatabeinghijackedbymaliciousactors.Atthesametime,consumersconcernedaboutstoringPersonalIdentifiableInformation(PII)inthecloudwillnolongerneedtoworry,assuchdatawillbeprocessedinedgedevicesandlocalservers.

PREPARINGFORTHEAIOFTOMORROW

WhenwecomparethestateofAItodaywiththevisionfortomorrow’sAI,afewdistinctcharacteristicsjumpout,asshowninFigure4.

Figure3:AIEvolutionandTechnologicalChallengesandRequirements

(Source:ABIResearch)

TechnologicalChallenges

Specializedandsiloed

AIofTodayAIofTomorrowandRequirements

MultimodallearningmodelwithpowerfulAIchipsetforreal-timeinference

Multi-taskandfullintegration

ContinualimprovementofNAS,modelcompression,andAutoMLSolution

Eithersmallorbig

Diverseinsizeandresourceequipment

Edgecloudoredge

Federatedlearningmodelsthatbenefit

fromincreasingdiversityofcomputing

nodesandhighdatabandwidth

Edge-to-cloudcontinuum

DataandAIdevelopmentplatform

thatcanexplainprediction,visualize

behavior,andresolvebiases

Blackbox

4

Transparentandexplainable

PREPARINGTHEFOUNDATIONFORTHEAIOFTOMORROW

TotrulymaketheAIoftomorrowareality,itisapparentthatbusinesseswillneedtocontinuetheirinvestmentintherightAIinfrastructureandapproach.

NewLearningTechniques

TheevolutionofAIwillbeshapedbytheemergenceofnewAItechniques,suchasfederated

learning,multimodallearning,graphDL,reinforcementlearning,andmeta-learning.TheseAI

techniquesuseafullydistributed,highlycustomizedcomputinglandscapetolearnandhandlemultipletasks.Theabilitytointerpretcontextualinformation,understandthelinkagebetweendifferentfactors,andworkandlearnfromotherAImodelscreatesimmenseopportunitiesforAItochangehowbusinessesoperate.Theywillbereadytohandlemorecomplicatedtasksthanhumanemployeescanhandle.

Table1:KeyAITrendsandTheirBusinessImplications

(Source:ABIResearch)

NewAI

Techniques

Descriptions

Strengths

TechnologyPlatformRequirements

Federatedlearning

FederatedlearningisadistributedMLapproachinwhichmultipleuserscollaborativelytrainamodelwithoutmovingdatatoasingleserverordata

center.Instead,eachcomputenodewillexecutethesamemodel,trainsuchamodelonthelocaldata,andthuscomputeandstorealocalversionofthemodelineachnode.

FederatedlearningprovidesedgedeviceswithqualityMLmodelswithoutcentralizingthedata.Deployedinvariousenvironments,includingsmartphones,healthcare,andfinance,thetechniqueallowstheaccessofdatasetsfromdifferentusers,institutions,ordatabases,whilehelpingtocomplywithrequiredprivacyandconfidentialitylaws.

Stableandubiquitous

connectivitybackbone,withAIcomputetakingplacein

localandcloudenvironments.Frequentexchangeofdataandsyncbetweenthecloudanddifferentedgenodestoensurethemodelisuptodate.

Multimodallearning

Multimodallearningcansimultaneouslyprocessvariousdatatypes(image,text,speech,numericaldata)usingmultiplealgorithms.MultimodalAIcaninterpretsuchmultimodalsignalstogetherandmakedecisionsbasedoncontextual

understanding.

AI-basedonmultimodallearningcanmimichumandecision-makingbyingestingdifferentdatasources.Asaresult,itoftenoutperformssingle-modalAIinmanyreal-worldproblems,suchascustomerservices,clientengagement,andpatientcare.

Databasesthatcaningest

variousdatasourcesandAImodelsthatprocessdifferentdatamodalitiesandperforminferenceinreal-time.

Reinforcementlearning

ReinforcementlearningisanMLtrainingmethodthatrewardsthelearningagentwhenmaking

desiredbehaviorsandpunishesitwhenmakingundesiredones.Generally,areinforcement

learningagentcanperceiveandinterpret

itsenvironment,takeactions,andlearnthe

associationsbetweenstimuli,activities,andtheoutcomesofitsactionsthroughtrialanderror.

Reinforcementlearninghasbeenwidelyadoptedin

simulationtotrainandretrainbehaviorsofautonomousvehiclesandrobotsfortrafficmanagement,material

handling,routeoptimization,andspacemanagement.Asidefromthephysicalsystem,thesoftwarecanalsobetrainedusingreinforcementlearningforadata-drivenproductorprocessoptimization,suchassupplychainoptimization,prototyping,andgenerativedesign.

PowerfulAIcomputeplatforminthecloudwithprecise

andrealisticrenderingof

thereal-worldenvironment.

Alternatively,ahighlyoptimized,unsupervisedself-learning

modelinenddevices.

GraphNeuralNetwork(GNN)

GNNsareDLneuralnetworksdesignedtoperforminferenceondatastoredingraphdatabases.Graphdatabasesconnectspecificdatapoints(nodes)andcreaterelationships(edges)intheformofgraphsthattheusercanthenpullwithqueries.TheAIcanunderstandtheinterdependencybetweeneach

datapointandprovidesrelevantpredictions.

GNNsareidealforanalyzingaspecificissuethat

involvesnumerousfactors.Forexample,creditriskforcreditcardcustomersrequiresunderstandingcurrentcreditscores,credithistory,employment,income,andothersocio-economicfactors.Otherusecasesincluderecommendationsystems,molecularcellstructurestudy,readingcomprehension,andsocialinfluenceprediction.

HardwareandsoftwareareoptimizedforGNNs,asGNNstakemuchlongertotrain.

Meta-learning

Meta-learningreferstoMLalgorithmsthat

learnfromtheoutputofotherAIalgorithms.

Thesemodelscanlearnacrossasuiteofrelatedpredictiontasksthroughanadaptiveprocess,

allowingthemtospeeduptheirlearningprocesswhilelearningmultiplefunctionssimultaneously.

Meta-learningisstillembryonic,soitisratherdifficult

topredicthowinfluentialandimpactfulthetechnologywillbecome.Nonetheless,ifthecurrentprediction

isaccurate,meta-learningmodelscanlearnquickly,

requiringlesstrainingtimeandfewerresourcestodesignanddevelop.Thiswillsaveenormousamountsoftimeandacceleratetimetomarket.

Ultra-high-performance

hardwareandsoftware,asmeta-learningisanensembleofthemostadvanced

MLtechniques,suchas

reinforcementlearningandtransferlearning.

5

PREPARINGTHEFOUNDATIONFORTHEAIOFTOMORROW

Moreimportantly,theyallowbusinessestoscaleoutandscaleupdependingonbusiness

needswithoutneedingtoexpandandtraintheirworkforce.Someofthesetechniques,suchasfederatedlearningandreinforcementlearning,havealreadybeenadoptedbycloudAIgiantsandlargecorporationstodesignadvancedAImodelsinlarge-scalerecommendersystems,

frauddetection,andvirtualassistance.Meta-learningandGNNs,ontheotherhand,areslowlymaturingandappearinginsomeinterestingusecases,suchasdrugdesign,roboticstraining,anddiseasediagnostics.

OptimizedAIInfrastructure

Inrecentyears,AItechnologyprovidershaveprogressivelyreducedthebarrierstoentryby

activelylaunchinginnovativeproductsandservices.GraphicProcessingUnits(GPUs)and

Application-SpecificIntegratedCircuits(ASICs)arebeingadoptedforAItrainingandinference.

Nowadays,moreandmoregeneral-purposeCentralProcessingUnits(CPUs)cansupportAI

inferenceandtraining.Theintroductionofpre-traininglanguagemodelsallowsdevelopersto

buildcomplexapplications,suchasspeechrecognitionandmachinetranslation,withouttrainingamodelfromscratch.AutoMachineLearning(AutoML)providesmethods,tools,andtechniquestomakethedevelopmentprocesseasierfornon-AIexpertsbyautomatingAIworkflow.

Notsurprisingly,alltheseadvancementshaveledtoahugedemandforAIchipsets.AccordingtoABIResearch’sArtificialIntelligenceandMachineLearningmarketdata(MD-AIML-109),theglobalAIchipsetmarketisestimatedtobeUS$32.3billionin2022.ThisincludesthesalesofAItrainingandinferencechipsets,includingtheCPU,GPU,FieldProgrammableGatedArray(FPGA),NeuralProcessingUnit(NPU),AIaccelerator,microcontroller,andneuromorphicchipset,inalldata

centersandenddevices.Furthermore,thedemocratizationofAIwillleadtoAIbeingdeployedacrossawiderangeofphysicalsitesandcomputenodes.Asaresult,thismarketisexpectedtogrowtoUS$68.8billionin2027,withaCompoundAnnualGrowthRate(CAGR)of26%.

Chart1:TotalRevenuefromAIChipsetSales

WorldMarkets:2020to2027

(Source:ABIResearch)

35

AnnualRevenue(US$Billions)

30

25

20

15

10

5

0

CloudAIChipsetsEdgeAIChipset

6

20202021202220232024202520262027

PREPARINGTHEFOUNDATIONFORTHEAIOFTOMORROW

KEYPRINCIPLESOFAIINFRASTRUCTUREINVESTMENT

Today’sAIisnarrowlyfocused,requiresawiderangeofexpertise,andexistsinasilo.Incontrast,theAIoftomorrowrequiresenormousamountsofresourcesanddeeptechnologicalknowledge,whichremainsoutofreachformostbusinesses.Therefore,businessesmuststartearly,identify

thebusinessoutcomesthatcannotbeeasilyachievedwithoutAI,activelybuildinternalcapabilities,androllouttheseadvancedAItechniqueswidelyacrosstheentireorganization.Insummary,

belowarefourkeypillarswhenconsideringAIinfrastructure:

?AIInfrastructureMustBeDrivenbyBusinessOutcomes:ThevisionofAIinfrastructuremustbebasedontheintendedbusinessoutcomeofAIdeployment.Businessesmustfirstunderstandtheshort-andlong-termvaluesAIbringstotheiroperationbeforedesigning

themostsuitableAImodels.WhenanAIprojecthasaclearbusinessoutcome,ithasactualfinancialvaluesthatseniormanagementcanrecognize.

?AIInfrastructureMustBeHeterogenousandFlexible:TounlocktheactualvalueofAI

andyieldmaximumbenefits,scale-upandscale-outofAIapplicationsarecritical.Building

anAIinfrastructurethatofferstheproperfoundationtosupportdifferentfacetsofAImodeldesign,development,anddeploymentacrossdifferentcomputingplatformsgoesalongwaytoprotectandfuture-proofcurrentinvestments.AheterogeneouscomputeplatformwillofferthebestperformanceacrossallAItasks.AIdeveloperscanusetheCPUfordatagatheringandpreparation,beforeswitchingtotheGPUandASICformodeltraining,andfinallyusingeithertheGPU,ASIC,orCPUforAIinferenceworkload.

?AIInfrastructureMustBeBackwardCompatible:AllAIinfrastructuremustbeableto

workwithexistingenterprisesolutions.Therefore,settingaversatile,robust,andinteroperablefoundationwithallexistingsolutionsisamust.Incompatibilityriskscreatingmanysilosinthebusinessoperation,leadingtopoorlyoptimizedIT/OTinfrastructureandprocesses.

?AIInfrastructureMustBeOpenandSecure:Businessesalwayswanttoavoidvendorlock-in.AnAIinfrastructureconsistingofopenhardwareandsoftwarethatcaninteroperatewithothersolutionsissignificantinensuringsmoothIT/OTprocesses.Atthesametime,opennessshouldnotleadtoacompromiseinsecurity.TheAIfoundationmustfeaturestate-of-the-artcybersecurityanddataprotectionmechanismstopreventhacking,protectuserdata,and

7

complywithlegalrequirements.

PREPARINGTHEFOUNDATIONFORTHEAIOFTOMORROW

ItisclearthatAIisstillinitsinfancy,andbuildingtheproperfoundationforitiscriticalforits

futuresuccess.InsteadoflookingatAIfromtoday’slens,allbusinessesmusthaveaclearlong-termplan.Thisvisionwillhelpthemnavigatethechallengesandtechnologyrequirementsfor

AI,helpingthemmaketherightdecisionininvestinginthemostoptimalandfuture-proofAI

infrastructure.ThefollowingsectiondiscussesvariousapproachesbusinessescantaketodeployAI.Inaddition,ithighlightsthekeyfeaturesandcharacteristicsbusinessesmustpayattentiontowhenselectingtheirAItechnologies.

CHARACTERISTICSOFAFUTURE-PROOFEDAIINFRASTRUCTURE

EnablingAIbroadlyacrossenterprisesrequiresadifferentmindset.BusinessesmustunderstandthatbuildingAIforbusinessisacontinuousprocessinvolvingmanybuildingblocks.AsshowninFigure5,severalkeyrecentadvancementshaveallowedAItobecomeareality.

Figure4:KeyAITechnologyTrends

(Source:ABIResearch)

User-FriendlyOpen-SourceAITools

DataScienceand

AIDeveloperCommunity

IT/OTConvergence

GrowingAIEcosystem

AI

Democratizationof

ComputingResourcesDemocratizationof

EdgeAI

EmergenceofLargeMLModels

8

AllofthemhaveasignificantimpactonhowbusinessesshoulddeploytheirAI.

PREPARINGTHEFOUNDATIONFORTHEAIOFTOMORROW

9

PREPARINGTHEFOUNDATIONFORTHEAIOFTOMORROW

Table2:KeyAITechnologyTrendsandTheirBusinessImplications

(Source:ABIResearch)

KeyTrends

Descriptions

BusinessImplications

IT/OT

convergence

ITandOThavetraditionallybeendevelopedseparately,withnoabilitytoexploitoperationsandproductiondatatomakemoreinformeddecisionsforoptimizedworkflowand

well-plannedproductionandmaintenanceprocesses.COVID-19hasspedupthedigitaltransformationprocessinmanybusinesses,leadingtoemergingtechnologies,suchastheInternetofThings(IoT)androboticsautomation.Asaresult,businessesalsostarttocollectmoreoperationaldatathatareveryusefulforplanning,optimization,andupgrades.

Thisconvergencewillenablebusinessestogaininsightandmakedata-drivendecisions.Theycanalsooptimizetheirexistingworkflowswithoutneedingtoscaleuprapidly.

Democratizationofcomputingresources

ThedemocratizationofcomputingresourcesforAIhasbeenachievedbythewideavailabilityofpubliccloudcomputing.Indeed,AIdevelopersmaywanttoleveragethecentralized

processingandstorageofferedbyCSPsinsteadofdeployingtheirownAIhardware,whichisnoteconomicalenough,mainlyforexecutingdenseAInetworks.Theuniformityand

scalabilityofCSPs’computeandstoragearchitectureenablethemtohandlecompute-intensiveDLmodelsonanon-demandbasis,significantlyloweringthebarrier-to-entryforAIdeveloperswithouttheabilitytobuildandmaintaintheirownAIinfrastructure.AsAImodelsarebecomingmorecomplex,theaccessprovidedbyCSPsisessentialforthedemocratizationofAI.

BusinessesmustinvestintherightAIinfrastructurebybuyingfromestablishedCSPsorbuildingtheirprivatecloudinfrastructure.Whilepubliccloud

solutionsarescalable,theycanbecostlywhen

comparedtobeingwell-plannedforprivate

infrastructurebasedonlong-termsgoals.Therefore,businessesshouldalsoconsiderleveragingthe

bestofbothworldswithhybridclouddeploymentstogetthebestprice-performanceadvantageandflexibility.

EmergenceoflargeDLmodels

Anotherprimaryreasonbehindthegaininaccuracyandperformanceisthegrowthin

DLmodels,preciselythenumberofparametersandhyperparameters.AImodelshave

scaledsignificantlyinthepastyears.Dependingontheapplication,largemodelsprovide

fundamentallyuniqueadvantages.Forexample,OpenAI’sGPT-3,widelyconsideredthemostadvancedNaturalLanguageProcessing(NLP)modelof2021,has125millionto175billionparametersandcanhandleadvancedapplications,suchasgenerativeemailsordocumentsummaries.NewermodelslikeBLOOMfromBigScience,whichhas176billionparameters,supportmultiplehumanlanguagesandprogramminglanguages.

BusinessesneedtoconsiderthecostofAItrainingandimplementation.Forcontext,thecloud

computingcostforthetrainingofBLOOM,whichisaround330Gigabytes(GB)insizeisestimatedtobeinthemulti-million-dollarrange.Asidefromtheproperhardwareinfrastructure,businessesmustalsoidentifythesuitableapplicationsandusecasestheywanttodeploy.

DemocratizationofedgeAI

HighlyoptimizedandminiaturizedAImodelsarecurrentlyembeddedinsmartsensors,

devices,andgateways.Thesecarefullycraftedsmallermodelscanalsoperformnarrowly-focusedapplications,specificallyinalways-oncomputervisionandtime-seriesdataanalysis.Solutionprovidershaveintroducedpower-efficientAIprocessors(ASIC,NPU,neuromorphicchipset),DLmodeloptimizationtechniques(knowledgedistillation,pruning,and

quantization),developer-friendlytoolsandservices,andmoreintelligentresourceallocation.

EdgeAIisawaytominimizelatency,privacyrisk,

andconnectivitycosts.Businessesmustlookintothelong-termbenefitsofedgeAIanddevelopastrategytodeployintheiroperation.

GrowingAIecosystem

TheAIecosystemcontinuestogrowatarapidrate.AIstartupsareofferingawiderangeofsolutions.Themostwell-knownstartupsworkonfacialrecognition,AdvancedDriver-AssistanceSystems(ADAS),andNLP.COVID-19hasbecomeacatalystbehindtherapidadoptionofAI-basedenterpriseautomation,suchasRPA,AI-aidedspeechrecognition,transcriptionandtranslation,andsalesandmarketingenablementtools.ThoselookingtobuildtheirowncustomAIwillleantowardstar