Abstract

Edgecomputing,asakeytechnologyofthenextgenerationofradioaccess

networks(RAN),hasdriventhedecentralizationofnetworksandcomputingfacilities.

Edgeserversclosertouserterminalscansignificantlyreduceservicelatencyandcope

withemergingnewscenarios.Simultaneously,therapiddevelopmentofartificial

intelligence(AI)playsasignificantroleinenhancingtheperformanceofedge

computing,aidingedgedevicesincopingwiththerapidlyincreasingdataontheedge.

Therefore,combiningthelocalcomputingcapabilityofedgedatawiththestrong

computingcapabilitiesofAI,knownasedgeintelligence,canenhancethedata

processingcapabilitiesontheedge,improvetheoverallperformanceofwireless

communicationsystems,andenhanceuserserviceexperiences.Edgeintelligenceisa

hotandrapidlydevelopingfieldinrecentyears,andthiswhitepaperaimstoanalyze

thecurrentresearchprogressinedgeintelligence.Itmainlyincludes:

(1)6GEdgeIntelligenceNetworksandInfrastructure:Firstly,theedge-native

intelligentarchitecturefor6Gnetworksisanalyzed.Then,theedgeintelligence

computinginfrastructureisintroduced,includingedgeintelligenthardwareandcloud

platforms.Finally,theedgeintelligencenetworkinfrastructureisdescribed,including

theedgeintelligenceaccessnetworkandcorenetwork.

(2)KeyTechnologiesofEdge-NativeIntelligence:Itisintroducedfromthe

aspectsofmodellightweight,edge-cloudcollaborativeintelligence,edgeintelligent

deployment,anddeepedgenodes.Edgeintelligenceinwirelessfederatedlearningis

alsoexplainedindetail,includingmodelsparsificationandmodelquantizationin

federatedlearning.

(3)ApplicationsofEdge-NativeIntelligence:Typicalapplicationsofedge-native

intelligenceareanalyzed,suchassmarttransportation,smartmanufacturing,and

smartenergysaving.

1/123

1.Introduction

1.1Background

From1Gto5G,communicationtechnologyhasundergonemultipleupgrades

andtransformations,significantlyimprovingdatatransferrates,reducinglatency,and

expandingnetworkcoverage.However,withtherapiddevelopmentoftechnologies

suchastheInternetofThings(IoT)andAI,theInternetofEverythingand

increasinglycomplexapplicationscenariosposechallengesthatexistingnetwork

architecturescannotmeet.Therefore,asthenextgenerationofcommunication

technology,6Gmustpossesshigherperformanceandmorepowerfulintelligent

capabilities,drivingthetransitionofedge-sidenetworksfrom"Internetof

Everything"to"IntelligenceofEverything."Tobetteradapttofuturediverseand

complexuserrequestsandapplicationscenarios,theconceptofedge-native

intelligencecameintobeingbyintegratingintelligenttechnologyintothedesignand

implementationofcommunicationsystems.[1]

Inrecentyears,thetheoryandtechnologyofAIhaveprogressedandfound

widespreadapplicationinindustrialscenarios.However,mostAIservicesare

typicallydeployedoncloudservers.Withtheadventofthe"InternetofEverything"

era,thenumberofterminaldevicesandtheamountofdatageneratedareincreasing

rapidly.Thecentralizeddataprocessingmethod,whichuploadsalldatatothecloud,

cannotmeetthelow-latencyrequirementsofusers.Consequently,edgecomputing

emergedwiththedevelopmentoftheInternetofThings(IoT)andAI.However,

currentresearchonedgecomputingimplementationfailstomeetthedemandsof

complexservicescenarios.Therefore,edge-nativeintelligencehasthepotentialto

becomethenextresearchhotspotinedgecomputing.[2]

Edge-nativeintelligenceenablesself-dynamicsensingandself-optimization

capabilitiesbetweenvariousunitsinthenativenetwork.Itbreaksawayfromthe

traditionalplug-inAIarchitecturebydeeplyintegratingAIintovariouslayersofthe

3/123

networktoenhancetheoverallsystemnetworkefficiency.Itachievesanautonomous

sensingoftheoveralllifecycleandself-managementwithinthenetwork

architecture.[3]

1.2OverviewofEdgeComputingandEdge-Native

Intelligence

Edgecomputing:Theconceptofedgecomputingisintroducedtoalleviatethe

processingpressureonclouddatacenters.Itisatechnologythatmigratescomputing

processesfromcentralserverstotheedgeofdevices.Thecoreideaistointegrate

network,computing,storage,andapplicationservicesintoaplatformclosetothedata

source,enablingservicestobeprovidednearby.Thistechnologyhelpsreducethe

processingloadofcloudcomputingandaddressestheissueofdatatransferlatency,

meetinguserneedsinreal-timeservice,intelligentapplications,security,andprivacy

protection.

Edge-nativeintelligence:Edgeintelligenceisthenextstageofdevelopmentafterthe

evolutionofedgecomputing.Withtherapiddevelopmentanditerationofedge

computingandAItechnologies,theconceptofedgeintelligencecameintobeing.It

executesAIalgorithmsattheedge,whichisamorecomplexdataanalysistask.

DeployingAIapplicationsonedgenodes,especiallyonmobiledevicesandIoT

devices,requiresthesupportofedgecomputing.Firstly,edgenodesneedtoprovide

correspondinghardwareandprogramminglibrariestomeetthebasicoperationsofAI.

Secondly,anedgecomputingplatformisneededforresourcemanagementandtask

schedulingonedgenodes.Finally,itisnecessarytosolvethetaskoffloadinganddata

securityproblemsincloud-basedcollaborativeAI.[4]

AsAItechnologycontinuestoevolve,thelevelofintelligenceinedgedevices

hasbeenelevated.Initially,edgeintelligenceprimarilyfocusedonrunningAI

algorithmsandmodelsonedgedevicestoachieverapiddataprocessingandresponse.

Thisapproachhadarelativelylowlevelofintelligencebecausethefunctionalityand

performanceofedgedeviceswerelimited,preventingtheexecutionofcomplexAI

4/123

algorithmsandmodels.[5]Withongoingtechnologicaladvancements,theperformance

andintelligenceofedgedeviceshavesignificantlyimproved.Inthisprocess,the

conceptofedge-nativeintelligencehasgraduallyemerged.Edge-nativeintelligence

emphasizestheintegrationofAItechnologyintoedgedevices,enablingthemtohave

autonomousdataprocessingandanalysiscapabilities.Thisapproachenablesedge

devicestobetteradapttocomplexapplicationscenarios,andimprovethespeedand

efficiencyofdataprocessingandresponse.[6]

1.3ImportanceofEdge-NativeIntelligence

Theimportanceofedge-nativeintelligenceincludesthefollowingaspects:

(1)FullUnleashingofDataPotentialattheNetworkEdgeThroughAI:Withthe

surgeinthenumberofmobiledevices,amassiveamountofdata(e.g.,audio,images,

andvideos)willbegeneratedatthedeviceend.TheintroductionofAIalgorithms

willbeessentialatthispoint,astheycanquicklyanalyzetheselargevolumesofdata

andextractfeaturesfromthem,leadingtohigh-qualitydecision-makingandimproved

efficiencyandreliabilityofdataprocessing.Thishelpstoreducemanualintervention

anderrorrates,improvingserviceefficiencyandreliability.[7]

(2)ExpansionofIntelligentAlgorithmDeploymentScopewithRicherDataand

ApplicationScenarios:Inthetraditionalcloudcomputingmodel,datasourcesare

generallyuploadedandstoredinthecloudduetoitsextremelyhighcomputing

performance.[8]However,withtherapiddevelopmentoftheInternetofEverythingera,

thetraditionalcloudcomputingmodelisgraduallyshiftingtowardstheedge

computingmodel.Inthefuture,theedgesidewillgenerateamassiveamountofIoT

data.IfallofthisdataneedstobeuploadedtothecloudforAIalgorithmprocessing,

itwilloccupyalargeamountofbandwidthresourcesandputagreatdealof

computingpressureonthecloudcomputingdatacenter.Toaddressthesechallenges,

offloadingcloudcomputingpowertotheedgeenableslow-latencydataprocessing,

thusachievingahigh-performanceedgeintelligenceprocessingmodel.[9]

(3)BetterSystemAvailabilityandScalabilitywithEdge-NativeIntelligence:AI

5/123

technologyhasachievedtremendoussuccessinmanydigitalproductsandservicesin

dailylife,suchasvideosurveillanceandsmarthomes.AIisalsoacriticaldriving

forceattheforefrontofinnovation,includingareaslikeautonomousdrivingand

smartfinance.Therefore,AIshouldbeclosertopeople,data,andterminaldevices.In

theprocessofachievingthesegoals,asdataprocessingoccurslocally,edgedevices

cancontinuetooperateevenifthecentralserverencountersissues.Additionally,with

theadditionofnewapplicationsorupgradestoexistingones,edgedevicescaneasily

expandormodify,providinggreaterflexibility.

(4)EnhancedAvailabilityandAccessibilityofAIApplications:Withthe

enhancedprocessingcapabilitiesofedgedevices,moreAIapplicationscanrunonthe

devicesthemselves,ratherthanrelyingsolelyoncloudservers.Thisincreasesthe

usabilityandaccessibilityofAI.[10]

References

[1]S.Talwar,N.Himayat,H.Nikopour,F.Xue,G.WuandV.Ilderem,“6G:

ConnectivityintheEraofDistributedIntelligence,”IEEECommunications

Magazine,vol.59,no.11,pp.45-50,Nov.2021.

[2]M.ElsayedandM.Erol-Kantarci,“AI-EnabledFutureWirelessNetworks:

Challenges,Opportunities,andOpenIssues,”IEEEVehicularTechnology

Magazine,vol.14,no.3,pp.70-77,Sep.2019.

[3]S.Deng,H.Zhao,W.Fang,J.Yin,S.DustdarandA.Y.Zomaya,“Edge

Intelligence:TheConfluenceofEdgeComputingandArtificialIntelligence,”

IEEEInternetofThingsJournal,vol.7,no.8,pp.7457-7469,Aug.2020.

[4]M.Pan,W.SuandY.Wang,“ReviewofResearchontheCurriculumfor

ArtificialIntelligenceandIndustrialAutomationbasedonEdgeComputing,”

2021InternationalConferenceonNetworkingandNetworkApplications(NaNA),

LijiangCity,China,2021,pp.222-226.

[5]Y.Xiao,G.Shi,Y.Li,W.SaadandH.V.Poor,“TowardSelf-LearningEdge

Intelligencein6G,”IEEECommunicationsMagazine,vol.58,no.12,pp.34-40,

Dec.2020..

[6]H.HuandC.Jiang,“EdgeIntelligence:ChallengesandOpportunities,”2020

InternationalConferenceonComputer,InformationandTelecommunication

Systems(CITS),Hangzhou,China,2020,pp.1-5.

[7]M.Mukherjee,R.Matam,C.X.Mavromoustakis,H.Jiang,G.MastorakisandM.

Guo,“IntelligentEdgeComputing:SecurityandPrivacyChallenges,”IEEE

CommunicationsMagazine,vol.58,no.9,pp.26-31,Sep.2020.

[8]Y.Sun,B.Xie,S.ZhouandZ.Niu,“MEET:Mobility-EnhancedEdge

inTelligenceforSmartandGreen6GNetworks,”IEEECommunications

Magazine,vol.61,no.1,pp.64-70,Jan.2023.

6/123

[9]Q.Cui,Z.Gong,W.Ni,Y.Hou,X.Chen,X.Tao,P.Zhang,“StochasticOnline

LearningforMobileEdgeComputing:LearningfromChanges,”IEEE

CommunicationsMagazine,vol.57,no.3,pp.63-69,Mar.2019.

[10]M.Yao,M.Sohul,V.MarojevicandJ.H.Reed,“ArtificialIntelligenceDefined

5GRadioAccessNetworks,”IEEECommunicationsMagazine,vol.57,no.3,pp.

14-20,Mar.2019.

2.6GEdgeIntelligenceNetworksand

Infrastructure

2.1Edge-NativeIntelligenceArchitecturefor6G

Asakeyenablingtechnologyforthenextgenerationofradiowirelessnetworks,

Multi-accessEdgeComputing(MEC)cansupportaplethoraofemergingservices.

WiththecontinuousdevelopmentofAI,itsapplicationinMECisbecoming

increasinglywidespread.However,in5Gnetworks,AIisonlyusedasanadd-on

applicationtoassistMEC.In6Gnetworks,MECwillincorporateAIfromtheinitial

designphase,treatingitasanintegralpartoftheMECsystem.Thisapproachaimsto

enhancetheflexibilityandopennessofMEC,betteraddressingtheconstantly

emergingapplicationscenariosanduserdemands.Asaresult,theedge-native

intelligencearchitecturehasbeenproposed,whichisbasedonthedecouplingand

reconstructionofAIfunctionstoprovideuserswithcustomizedAIservices.

2.1.1OverviewoftheArchitecture

Theedge-nativeintelligencearchitectureconsistsof"fourlayersandthree

planes",asshowninFigure2.1.The"fourlayers"includetheinfrastructurelayer,

virtualizationlayer,functionlayer,andapplicationlayer;the"threeplanes"include

thecontrolplane,AIplane,andmanagementandorchestration(MANO)plane.

7/123

Figure2.1Edge-NativeIntelligenceArchitecture

I.Fourlayers:

Infrastructurelayer:Locatedatthebottomoftheedge-nativeintelligence

architecture,itencompassesallcommunication,storage,andcomputingresources

inthesystem.CommunicationresourcesincludeWi-FiandtheInternet;storage

resourcesincludememory,HardDiskDrive(HDD)andSolidStateDrive(SSD);

computingresourcesincludeCentralProcessingUnit(CPU)andGraphics

ProcessingUnit(GPU).

Virtualizationlayer:Positionedabovetheinfrastructurelayer,itabstractsthe

underlyingresourcesintoaresourcepoolforusebyupper-layernetwork

functions.Whenservicedemandsarise,thevirtualizationlayercancreateDocker

containersandrunthemintheresourcepooltosupplynetworkfunctions,

ensuringtheirnormaloperationandtherebyguaranteeingcustomizedAIservices.

Functionlayer:Locatedabovethevirtualizationlayer,itconsistsofdecoupled

networkfunctions,namelycontrolfunctionsandAIfunctions,andaservicebus.

Differentnetworkfunctionscanbeactivated,released,andreconfiguredinreal

timebasedonservicerequirements,interconnectedthroughtheservicebus.

Applicationlayer:Locatedatthetopoftheedge-nativeintelligencearchitecture,

itincludesdiversenetworkapplications.Theapplicationlayerinteractsdirectly

withusersand,uponuserrequests,automaticallyinvokesthenetworkfunctions

8/123

ofthefunctionlayerandtheDockercontainersofthevirtualizationlayerto

provideservicestousers.

II.Threeplanes:

Controlplane:Itisresponsibleforthetransmissionandprocessingofcontrol

signalingfromtheinfrastructurelayertotheapplicationlayer.

MANOplane:IttransformsservicerequestsfromthecontrolplaneintoMANO

commandsandcoordinatesandmanagesthesystem'sfunctionsandresources.

TheMANOplaneincludestheVirtualizedInfrastructureManager(VIM),

FunctionalMANO,andApplicationMANO,dedicatedtothemanagementand

orchestrationofresources,functions,andapplications,respectively.

AIplane:AlsoknownasthenativeAIplane,itservesasthecoreaspectofthe

edge-nativeintelligencearchitecture,responsibleforlearninguserandnetwork

behavioranddemands,achievingself-operationofthenetwork.Itsvirtualization

layerprovidesaruntimeenvironmentlibraryforAIapplications,suchasPyTorch

andTensorFlow,whichcanbeselectedbasedonapplicationrequestsand

resourcestate.TheAIplaneincludesdecoupledAIfunctionsandaservicebusin

itsvirtualizationlayer,whileitsapplicationlayercomprisesatemplateselector

andanintelligentalgorithmmodellibraryforflexiblereconstructionof

edge-nativeintelligence.

2.1.2DesignandImplementationoftheNativeAIPlane

Intheedge-nativeintelligencearchitecture,themicroservice-basedAIplaneis

decoupledintoindependentAIfunctions.TheseAIfunctionscanbeactivatedand

invokedondemand.Whenanapplicationrequestarrives,thedecoupledAIfunctions

canbecombinedondemandtoprovideAIservicestousers,thusachieving

edge-nativeintelligence.

I.Decouplingofedge-nativeintelligenceplane:

AsshowninFigure2.1,intheedge-nativeintelligenceplane,AIservicesare

decoupledintoDataCollectionFunction(DCF),DataPreprocessingFunction(DPF),

9/123

ModelTrainingFunction(MTF),ModelValidationFunction(MVF),andData

StorageFunction(DSF).Eachfunctionisdescribedasfollows:

DCF:CollectsrawdatarequiredforAImodeltrainingandgeneratesthe

correspondingtrainingdataset.

DPF:Preprocessestherawdatacontaininginvalidcomponents.Removesinvalid

oroffsetcontentfromtherawdatathroughdatasampling,featureextraction,and

dimensionalityreduction.ConvertsthedataintotheformatrequiredforAImodel

training.

MTF:SelectstheappropriateAIalgorithmaccordingtoservicerequirementsand

trainsthecoremodeloftheAIalgorithm.

MVF:EvaluatestheperformanceoftheAImodelduringmodeltrainingor

real-timeinference.

DSF:StoresandmanagesalldataandAIF-relatedparametersoftheAIplane.

CommunicationandinteractionbetweendifferentAIfunctionsoccurthrougha

unifiedservicebus.Additionally,AIfunctionscancommunicatewithcontrol

functionsviatheservicebusandbeactivatedbyFunctionalMANObasedonservice

types.

II.Reconstructionofedge-nativeintelligenceplane:

Edge-nativeintelligencereconstructionborrowstheideaoftemplateand

instantiation.ItperformsAIfunctionactivation,runtimeconfiguration,andresource

allocationbasedonservicetypetoachievecustomizedAIservices.

Template:Providesacommonsolutionforaclassofedgeintelligentservicesby

extractingandabstractingtheircommonalities.Theedge-nativeintelligence

templateinvolveskeyelementssuchastemplateinformation(Tinf)andtemplate

identifier(Tid).TemplateinformationencompassesthecomponentsoftheAI

application,namelythetypesofAIF,requiredresources,andruntime

environments,storedintheintelligentalgorithmmodellibrary.Thetemplate

identifierdistinguishesdifferenttemplatescorrespondingtoAIapplicationsandis

storedinthetemplateselector.Beforeusingthetemplate,predefinedoperations

arenecessary,definingparametersrelatedtofunctionalityactivation,resource

10/123

allocation,andruntimeenvironmentconfigurationaccordingtospecificAI

applicationrequirements.

Instantiation:CreatesanAIapplicationinstancebasedontheparameters

definedinthetemplatetorespondtoAIservicerequests.AsshowninFigure2.2,

theedge-nativeintelligenceinstantiationprocessincludesthefollowingsteps:

1)MANOcontinuouslymonitorstheapplicationlayerandsendsatemplate

selectionrequesttothetemplateselectorwhenanapplicationrequestisreceived.

2)Thetemplateselectorselectsthecorrespondingtemplateaccordingtothe

applicationtypeandsendsitsTidtotheintelligentalgorithmmodellibraryto

requestTinf.

3)TheintelligentalgorithmmodellibraryextractsthecorrespondingTinf

ofthetemplateandprovidesfeedbacktothetemplateselector.

4)ThetemplateselectorsendsthereceivedTinftotheMANOplane.

5)TheMANOplaneperformstheinstantiationoperationaccordingtothe

receivedTinf:

(a)Configurestheruntimeenvironmentlibraryrequiredbytheapplication.

(b)Allocatestherequiredresources.

(c)ActivatestherelevantAIF.

11/123

Figure2.2Edge-NativeIntelligenceInstantiationProcess

2.2EdgeIntelligenceComputingInfrastructure

2.2.1EdgeIntelligentHardware

Withtherapiddevelopmentoftechnology,edgeintelligenthardwarehas

graduallybecomeafocalpointwheretheIoT,AI,andcloudcomputingintersect.This

typeofintelligenthardwarenotonlypossessesreal-timeandefficientdataprocessing

capabilitiesbutalsocanmakeintelligentdecisionsatthenetworkedge,significantly

alleviatingdataprocessingpressureonthecloudandimprovingoverallsystem

responsivenessandefficiency.

Intermsofcustomerdemands,edgeintelligenthardwarecaterstovarious

industries,placinghighrequirementsonadaptabilitytotheenvironment,real-time

12/123

processing,security,andstability.Forexample,insmartmanufacturing,edge

intelligenthardwarecancollect,process,andanalyzevariousdataonfactory

productionlinesinreal-time,enablingautomationandintelligenceintheproduction

process.Inthemedicalfield,edgeintelligenthardwarecananalyzepatients'

physiologicaldata,enablingremotehealthcareandintelligentdiagnosis.

Fromatechnicalperspective,edgeintelligenthardwareincorporatesadvanced

algorithmsanddataprocessingtechnologies,enablinghigh-efficiencydataprocessing

andanalysis.Additionally,itadoptsamultitudeofsensors,communication

technologies,andsoftwaredefinitions,achievinginterconnectednessand

interoperabilitywithvariousdevicesandsystems.Moreover,edgeintelligent

hardwarestandsoutwithitslowpowerconsumptionandhighreliability,readily

meetingtheusagerequirementsindiverseharshenvironments.

Intermsofproductforms,edgeintelligenthardwarecanmanifestinvarious

devicessuchasintelligentcameras,intelligentsensors,intelligentrobots,andedge

servers.Thesedevicescanconnectwithvariousequipmentandsystems,facilitating

datasharingandcollaborativeprocessing.Moreover,theycanundergoremote

managementandcontrolthroughthecloud,enablingremotemonitoringand

maintenanceofdevices.

I.Edgeintelligenthardwarerequirements

Asshowninthetablebelow,consideringthedistancefromthehardware

deploymentlocationtothedatacenter,edgeintelligenthardwarecanbecategorized

intoNearEdgeandFarEdge.NearEdgeprimarilyinvolvesthedescentofcloud

computing,resemblingclouddatacentersinfunctionality,withpowerfuland

comprehensivecomputingcapabilities.Thehardwareproductformsinclude

integratedcabinetsandheavy-edgeservers.FarEdgefocusesmoreonspecific

applicationsattheedgesite,withstrongrelevancetospecificapplicationssuchas

dataaggregation/transformations,protocolparsing,industrialcontrol,andAI

inference.Thehardwareproductformsarediverse,includingindustrialcomputers,

PLCs,gateways,andMEC.

13/123

FunctionProductExamples

Regionaldatacenters,CDN(contentdelivery

Deepedgecomputingnetworks),telecomdatacenters,hostingservice

Near

providers

Edge

Localdatacenters,heavy-edgeservers,microdata

Deepedgecomputing

centers(integratedcabinets)

Aggregationanalysisandcontrol,dataAIBox,MEC,HCI(hyper-converged

managementinfrastructure)

Aggregation,conversion,filtering,data

FarGateways,smallcells,routers,accesspoints

reduction,forwarding

Edge

Analogtodigitalconversion(sensors),

Industrialcomputers,PLC(programmablelogic

sendingcontroldata(actuators),direct

controller),DCS(distributedcontroller),etc.

analysis/control

Edgecomputinghardwareproductshavetheiruniquecharacteristics,distinct

fromthehardwareproductsofcloudcomputingandedgecomputing.Thereasons

behindthisdistinctionaretheprimarydemandsfacedbyedgecomputing:

（Ⅰ）Diverseandcomplexapplicationscenarios:

(1)Thediversityinedgedeploymentrequiresdifferentinfrastructure

combinations.Edgedeploymentspansvariousindustryapplications,userscenarios,

andverticaldomains.Itincludesawiderangeofinfrastructuresolutions,makingthe

edgesolutionecosystemhighlycomplexintermsofproductforms,configurations,

andmanagementtools.

(2)Edgecomputingisexperiencingrapidgrowthinindustriessuchas

telecommunications,utilities,manufacturing,andfinance.Telecomoperatorsare

activelybuildingedgecomputingplatforms,leadingmarketdevelopment.Other

industries,particularlyutilities,manufacturing,andfinance,arealsoacceleratingthe

adoptionofedgecomputingbydeployingdedicatededgeinfrastructuretoenhance

14/123

efficiencyinusecasessuchastheIndustrialInternet,gridmanagement,andsmart

commercialbuildings.

(3)ThevigorousdevelopmentoftechnologieslikeAI,machinelearning,big

datamodels,andheterogeneouscomputingfurtherpropelsthegrowthoftheedge

computingmarket.Theproliferationofcompute-intensiveanalyticalworkloadsis

ubiquitousinmanyindustriesandusecases,unlockingthepotentialofuntappeddata,

mostofwhichresidesorisgeneratedattheedge.Theexpectedconvergenceof

AI-nativecomputingcapabilitieswiththeperformancerequirementsofnewanalytical

platformswilldrivethegrowthofmanynewedgeinfrastructuredeployments.The

diversityofAIapplicationsalsodiversifiesthedemandforedgecomputinghardware,

software,services,andsolutions.

（Ⅱ）Longlifecycleproductdemands:

(1)Inedgecomputingapplicationsacrossvariousindustriesliketransportation,

healthcare,energy,andindustry,suchasrailtrafficcontrolsystems,mediumtolarge

medicalequipment,substation/distributionstationcontrolunits,andindustrialcontrol

DCS/MES,theproductsoftengothroughalonglifecycleinvolvingstagesofproduct

design,researchanddevelopment,testingandverification,implementationand

operation,andlatermaintenance.Therefore,5-7yearsorevenlongerlifecyclefor

edgecomputingproductsiscrucialfortheseapplicationscenarios.Thisnotonly

implieshigherstabilitybutalsomeanslowermaintenancecostsinthelaterstages.

(2)Thelonglifecycletargetstheentireservicesystem,includingnotonlyedge

computinghardwaredevicesbutalsoplatforms,serviceapplications,protocols,and

generateddatarunningonthehardwaredevices.

(3)Theextendedlifecycleencompassesnotjusttheruntimebutalsotheongoing

provision,service,updatestohardwaredevices,andthecontinuousevolutionof

platformsandserviceapplications.

（Ⅲ）Demandingoperatingenvironments:

(1)Harshphysicalconditions

Edgedevicesaredeployedindiverselocations,facingcomplexphysical

15/123

environ