TechnicalWhitePaper

MobileNetworks

EvolutionintheAIEra

September2025

SamsungNetworks

Contents

1.Introduction 4

2.Software-basedFlexibleNetwork 6

2.1.NetworkEvolution:fromPurpose-builtHardwaretoSoftware-basedNetwork

2.2.FlexibilityofSoftware-basedNetworks

3.IntegrationofAIandMobileNetworks 12

3.1.WhySoftware-basedNetworksarebestsuitedforAI

3.2.AI-poweredAutomation

3.3.AI-poweredCoreandRadioAccessNetwork

4.PersonalizedService-centricNetwork 18

4.1.PersonalizedNetwork-as-a-Service

4.2.AI-poweredAutonomousServiceAssurance

4.3.On-DemandEdgeAI-as-a-Service

5.Conclusion 22

2

3

Abbreviations

AI

ArtificialIntelligence

mmWave

Millimeterwave

API

ApplicationProgrammingInterface

MNO

MobileNetworkOperator

CAPEX

CapitalExpenditure

NaaS

Network-as-a-Service

COTS

Commercialoff-the-shelf

NF

NetworkFunction

CPU

CentralProcessingUnit

OPEX

OperationalExpenditure

CSR

CellSiteRouter

RAN

RadioAccessNetwork

C-RAN

CentralizedRadioAccessNetwork

SLA

ServiceLevelAgreement

CU

CentralizedUnit

SoC

Systemon-Chip

D-RAN

DistributedRadioAccessNetwork

UE

UserEquipment

DU

DistributedUnit

UPF

UserPlaneFunction

E2E

EndtoEnd

vCore

virtualizedCorenetwork

GPU

GraphicsProcessingUnit

vCSR

virtualizedCellSiteRouter

KPI

KeyPerformanceIndicator

vDU

virtualizedDigitalUnit

LLM

LargeLanguageModel

vRAN

virtualizedRadioAccessNetwork

MIMO

MultipleInput,MultipleOutput

4

1.Introduction

Sinceitsinceptioninthe1970s,themobileindustryhasundergonearemarkabletransformationdrivenbytheaspirationtotransmitmoredatainlesstime.Theevolutionbeganwithfirst-generation(1G)networks,whichofferedonlyafewkilobitspersecondofvoice.Witheachsubsequentgeneration,thecapabilitiesexpandedexponentially:2Gusheredinspeedsuptohundredsofkilobitspersecond,while3Gpeakedatseveralmegabitspersecond.Theadventof4Grevolutionizedaccess,supportinghigh-speeddatatransmissionofupto1Gbps.Now,withthearrivalof5G,userscanexperiencelightning-fastservicesexceeding10Gbps.Thisleapintechnologyenableshigh-definitionvideostreaming,immersivemobilegaming,andseamlessmobileofficeexperiencesanytimeandanywhere.

Whatdoesthefutureholdforthemobileindustry?

Samsungenvisionsagroundbreakingevolutioninmobilenetworksthatwillfurtherenhancetransmissioncapacity.Thesenetworkswillbecharacterizedbytheirflexibility,automation,intelligence,andhaveastrongfocusondeliveringservice-orientedsolutions.Thisprogressiveapproachpromisestorevolutionizehowweconnectandinteractinanincreasinglydigitalworld.

1。

Thefutureofmobilenetworksliesintranscendingtraditionaldedicatedhardwaresetupsandembracingatransformativeshifttowardhardware-agnosticsoftwaresolutions.Bydecouplinghardwarefromsoftware,mobilenetworkoperators(MNOs)canleveragecommercialoff-the-shelf(COTS)orgeneral-purposehardware,enablingthemtoinstallandrunnetworksystemsseamlesslywithouttheneedforhardwaremodifications.ManyMNOsarealreadyleveragingsoftware-drivencoreandRANequipment,andthispromisingtrendwillundoubtedlygainmomentumwithitsstrengths.Intheverynearfuture,amajorityofmobileinfrastructure,includingthecorenetwork,RAN,andtransportnetworkequipment,willtransitiontoasoftware-centricmodel,drivinginnovationandefficiencyacrosstheindustry.

2。

Mobilenetworksareonthebrinkofatransformativeshifttowardsautomation,significantlyreducingtheneedforhumanintervention.Aswefaceincreasingcomplexityinmobilenetworks—drivenbyevolvingcapacitydemandsandfunctionaladvancements,thesediversecomponentsmustworkharmoniously.Networkautomationstreamlineseveryphase,frominitialinstallationandongoingoperationstopracticalproblemanalysisandresolution.Thismethodologyreducescostsandminimizesthelikelihoodofhumanerrors,resultinginenhancednetworkqualityandstability.Embracingthisautomatedapproachiscrucialforthefutureofmobilenetworks.

5

3

Mobilenetworkswillbecomemoreintelligentthroughtheapplicationofartificialintelligence(AI),takingovertaskstraditionallyperformedbyhumansortasksrequiringhumanexpertise.Forinstance,withlargelanguagemodels(LLMs)andintention-basedautomation,networkscandelivermoreefficientservicesbasedondata-driveninsights.Additionally,thesesystemswillfacilitatefasterandmoreaccuratedecision-makingbyproposingoptimizedsolutionsforhumanconsideration.Withadvancedlearningcapabilities,AIcanminimizenetworkdowntimeandenhanceservicecontinuitybyeffectivelyidentifyingtherootcausesofproblemswhentheyarise.

4。

Mobilenetworksandproductswillbetailoredtobeservice-orientedandpersonalized.Software-basedflexiblenetworksenablethecreationofcustomizedapplicationsthatsatisfymorespecificneeds.WhenpairedwithautonomousnetworksandinnovationsdrivenbyAI,thesesystemsenhanceresourceallocation,predictusagepatterns,andoptimizeservicelevelagreements(SLAs).Thiscombinationsystematicallyimprovesservicequalityandensuresanexcellentenduserexperience.

TheevolvinglandscapeandshiftsalsopresentnewbusinessopportunitiesforMNOs.Thesemobileinfrastructures,whereallnetworkequipmentoperatesthroughsoftware,relyongeneral-purposecentralprocessingunits(CPUs)andgraphicsprocessingunits(GPUs).Samsungenvisionsthisinnovativeinfrastructureasessentiallytransformingmobilenetworksintoadatacenterandcloudplatformcapableofhandlingavarietyofworkloads,includingvariousAIapplications.Mobilecommunicationequipmentislocatedwhereverusersare,fromcentraldatacenterstoantennasites,ensuringservicesaredeliveredclosetousers.Byleveragingfuturenetworks,MNOscanunlocknewrevenuestreamsnotonlyfrommobilecommunicationservicesbutalsofromcloudoperations[1].

6

2.Software-basedFlexibleNetworks

Theevolutionofmobilecommunicationtechnologyhasenabledittomeettheincreasinguserdemandfornewservices–andtheseadvancementswouldn’tbepossiblewithoutchipinnovations.Continuousdevelopmentsinchiptechnologyhaveintroducedmicro-processing,whichenabledtheimplementationofincreasinglycomplexandadvancedtechnologiesatlowercosts,smallersizes,andreducedpowerconsumption.Smartphonesreleasedin2010wereequippedwithchipsetsmadeusing65nm(nanometers)processtechnology,whilesmartphonesreleasedin2025utilizechipsetsbasedon3nmprocesstechnology,enhancingperformanceanduserexperience.Mobilenetworkequipmentalsohasapurpose-builtsystem-on-chip(SoC)forprocessinglarge-capacitydata,andequipmentprovidersareconsistentlyintegratingthelatestSoCsintotheirbasestations.

However,costsassociatedwithdevelopingnewchipshavegrownexponentially,particularlyforhigh-techprocesses.Forinstance,itcost$48milliontocreatea28nmchipin2011,whereasthedevelopmentcostfora2nmchipnowstandsat$725million.Thisevolutionrepresentsastaggering15-foldincreaseinchipdevelopmentcostsoverthepastdecade(seereferenceFigure1),whichpromptedacallforchangesinthedevelopmentofmobilecommunicationequipment.

Moreover,thereleasecycleforpurpose-builtSoCsproducedbymobileequipmentcompaniestypicallyrangesfromthreetofouryears,notablylongerthantheonetotwo-yearcycleforgeneral-purposeCPUsandGPUs.Asaresult,itisunclearwhetherchipsdesignedforspecificpurposeswillremaincompetitiveorcontinuetobeutilizedinthenearfuture.Incontrast,general-purposechipsalreadyholdsignificantmarketshare,havearobustecosystem,andarerapidlyadvancing.Theymayserveasamoreeffectivealternativeinthemobilecommunicationssector.

source:ArmInc2023IPOprospectus(IBS,july2022)

$725M

]

designcost[USD

$48M

processnode

Figure1.Chipsetmarkettrend–chipdesigncostperprocessnode

Thatsaid,operatingsoftware-basedmobileequipmentongenericchip-basedhardwareallowsforsubstantialcostsavingscomparedtodevelopingdedicatedchips.Thereareanumberofotheradvantagesthataredrivingthetransitiontosoftware-basedsolutionsongeneral-purposechipplatforms:

COTSequipmentcanbeutilizedregardlessofthehardwarebrandorunderlyingarchitecturetype.

Hardwareresourcescanbeusedmoreefficiently,ascapacitycanbeeasilyexpandedorreducedaccordingtomarketdemands.

Itenablesresourcesharingwithgeneralapplications,suchasAI,whichhelpstominimizeredundantinvestments.

7

2.1

NetworkEvolution:

FromPurpose-builtHardwaretoSoftware-basedNetwork

Effortsareunderwaytotransitionexistingcommunicationequipmenttowardasoftware-basednetworkacrossalldomains,RAN,core,andtransport.

Youcancheckthevideobyclickingontheimage.

01VirtualizedCoreNetwork

Corenetworkequipmenthastraditionallyhandledlarge-capacity,high-speedpacketprocessingusingdedicatedhardwarespecificallydesignedfortheuniqueattributesofeachnetworkentity.However,withtheperformanceimprovementsobservedwithgeneral-purposeCPUsandservers,itisnowpossibletovirtualizeandimplementcorenetworkentities.Thisshiftallowsforhighprocessingcapacitywithlowerpowerconsumption.

The5Gstandardintroducesseveral

changesaimedatenablingthe

virtualizationofthecorenetworkand

enhancingefficiency:

Theconnectionmethod

betweendifferentNFshasbeen

standardizedthroughaservice-

basedinterfacethatutilizes

HTTP/2

,allowingforeasieruseand

thereuseofservicesforeachNF.

Userequipment(UE)context

cannowbestoredinaseparate

databaseratherthanmanaged

individuallybyeachNF.Thisdesign

enablesallNFstooperatewithout

maintainingaconsistentstate,as

theycanretrieveandupdateUE

contextsthroughthedatabase.

Byimplementingthenetwork

entitieswithinthecorenetwork

domain-definedasnetwork

functions(NFs)–asmicroservices

inacloud-nativearchitecture,they

canbescaledinoroutbasedon

theCPUloadorthenumberofuser

sessionsbeingsupported.

Asaresultofthesetransformations,theprocessesandstorageforNFsaredecoupled,leadingtosimpleroperationsandimprovedscalability.

8

02VirtualizationofRAN

Traditionally,theRANreliedon

purpose-builtchipsduetothe

complexsignalprocessing

requirementsatthephysical

layer,whichinvolvesophisticated

timingandhigh-speeddata

processing.Dedicatedhardware

wasperceivedastheonlycost-

effectiveandenergy-efficient

solutionfortheseoperations.

Whilethatmayhavebeen

truepreviously,withtherapid

advancementsingeneral-purpose

CPUsandongoingeffortstoimplement

virtualizedRANs(vRAN),ithasbecome

feasibletovirtualizeRANequipmentto

competewithtraditionaldedicated

hardwaresolutions[2-4].

Inparticular,vRANoffersoptimaloptionsforvariousinstallationenvironmentsandrequirements,thankstoitsflexibleandscalableattributes,withouttheneedforadditionalormodifiedhardware.

Varioustypesofradiounits:asmultipleinput,

multipleoutput(MIMO)technologyevolves,multiple

typesofradiounits,rangingfrom2T2Rradiounitsto

64T64RmassiveMIMOunits,canbeconnected.

Variousversionsof3GPPstandards:eachMNOmust

support4GLTEand5GNR,dependingontheservice

status,whilealsomaintaining2GGSMformachine-to-

machineservices.Furthermore,itshouldbepossible

tominimizetheneedforhardwarechangesevenas

technologyevolvestoward6Ginthefuture.

Otherconfigurationoptions:in5G,RANfunctions

aredividedintothreecomponents:centralizedunit

(CU),distributedunit(DU),andradiounit(RU).With

theexceptionoftheRUs,whichrequireseparate

hardwareforhandlingradiofrequencies,each

distributedRAN(D-RAN)canincludebothCUandDU.

Alternatively,aD-RANmayconsistonlyofaDU,with

theCUcentralizedinaseparatedatacenter.

Variouscombinationsoffrequencybands:recent

cellularnetworksarerequiredtosupportavariety

ofbandwidthsandfrequencybands,dependingon

theMNOsandlocalenvironmentalconditions.These

mayincludenarrowbandwidthatFDDlow-band,

widebandwidthatTDDmid-band,andextremely

widebandwidthatmillimeterwave(mmWave).

Additionally,multiplecarriersmayneedtobe

supportedwithineachband.

ThesevariouscombinationsofRANrequirementscanleadtoeitherexcessorinsufficientcapacityifonlylimitedhardwareoptionsareused.Incontrast,vRANcanutilizeavarietyofcommerciallyavailableCOTSservers,allowingforoptimalconfigurationsthroughvRANsoftwareoptimizationtailoredtomeetcomplexneeds.Additionally,morepowerfulandcapablegeneral-purposeCPUsdesignedforcommunicationnetworksareexpectedtobereleasedinthefuture.Asaresult,vRANisanticipatedtocontinuetoadvancefurther,improvingitsperformance.

9

03DisaggregatedTransportNetwork

AsothercommunicationequipmentevolvesintovirtualizedsoftwareonCOTSservers,thereisagrowingmomentumtoadoptsystemsthatseparatehardwareandsoftwareintransportnetworks.

IncloudorMNOdatacenters,large-capacityroutersarenecessarytoconnecthigh-capacityserversandcoreequipmenttothepublicnetworkandD-RANsites.Theselarge-capacityroutersrequirenumerousphysicalEthernetportsandhigh-performancepacketprocessingcapabilitiestohandlesignificantnetworktrafficsimultaneously.

Consequently,itispracticallychallengingtoreplacethemsolelywithCOTSservers.

Instead,whiteboxrouterscanbeusedasdevicesthatmeetthesespecifications.Awhiteboxroutercombinesstandardnetworkingchipsforhigh-performancepacketprocessingwithageneral-purposeCPUforsoftwareoperation,effectivelyseparatinghardwarefromsoftware.Thisdesignallowsnetworkmanagerstocontrol,manage,automate,andoptimizetheentiretransportnetworkfromacentralizedlocationusinganopennetworkoperatingsystemandasoftware-definednetworkwhilesupportinglarge-capacitypacketswitching.

Youcancheckthevideobyclickingontheimage.

Anothertransportnetworkdeviceusedinmobilenetworksisthecellsiterouter(CSR)attheD-RANsite.Unliketraditionallarge-capacityrouters,D-RANsitesdonotrequireahighnumberofEthernetportsorextensivesimultaneousnetworkswitchingprocessing.Asaresult,avirtualizedCSR(vCSR)canbeimplementedusingcomputingresourcesfromaCOTSserverwhereavRANhasalreadybeeninstalled.TheintroductionofavCSRattheD-RANsiteeliminatestheneedforaseparatehardwareCSR,integratingallfunctionsintoasinglebox.ThisintegrationisbasedonthecapacityrequirementsoftheD-RANsiteandadvancementsingeneral-purposeCPUtechnology,leadingtosignificantsavingsinbothcapitalexpenditures(CAPEX)andoperationalexpenditures(OPEX)[5].

10

2.2FlexibilityofSoftware-basedNetworks

0102

VirtualizationofCoreand

TransportNetworksandtheRoleofEdgeComputing

Thecoredomainfunctions—includingthecontrolplanefunctionandtheuserplanefunction(UPF)forprocessinghigh-capacitycallsfromeachRANsite—arelocatedintheMNO’scentraldatacenter.ThisdatacenteralsomanagesRANsitesandhandlesapplicationprocessingforservicesprovideddirectlybythespecificMNO.VirtualizingNFsinthecentraldatacenteroffersseveraladvantages,includingdynamicresourceallocationandadjustmentinresponsetochangesinsubscribernumbersortrafficvolume.ThisflexibilityenablesthenetworktoadaptefficientlytonewRANsitesorservices.

FutureNetworksandIntegrationofAI

Whenwethinkofafuturenetworkwhereservicesarecontinuouslyevolving,edgeapplicationswillincorporateAIprocessing.IftheAIoperationsaresensitivetolatency,theAIprocessescanbeaddedtothevirtualNFwithintheedgedatacenter.Alternatively,iftheAIoperationsrequiresubstantiallearningfromdiverseandextensivedatacollectedfrommultipleedgedatacenters,theAIprocessesmightbehousedasavirtualNFinthecentraldatacenter.Inthiscase,thecomputingresourcesintheedgedatacentercanberepurposedforotherfunctions.Asthecomputingresourcesoftheedgedatacenterevolveandexpandinvariousways,thisdevelopmentpresentsMNOswithanopportunitytoestablisha“distributededgecloud.”

central

datacenter

platformn

Edge

platform

datacenter

C-RAN

D-RAN

datacenter

datacenter

platform

Figure2.Software-basednetworkfunctionsandAIworkloadsrunningongeneral-purposeplatforms

11

03

ReducingLatencywithEdgeDataCenters

Shorteningthetrafficpathbetweenusersandapplicationserversisessentialtosupportlatency-sensitiveservices.Thisroutingcanbeachievedbyrelocatingapplicationserverstoedgedatacentersclosertousers.Insteadofroutingtrafficthroughthecentraldatacenterofanationalnetwork,theedgedatacentercanbeintroducedtosubstantiallyreducethedistancebetweenusersandserviceapplications.IftheRANsitesconnectedtotheedgedatacenterconsistsolelyoftheDUandtheRU,theedgedatacentermustincludeseveralnecessaryNFs.TheseincludetheCUoftheRANdomainandtheUPFofthecoredomain.

Sincethe5GstandarddefinestheCPandUPofthecoredomainseparately,theedgedatacentercanconnectwiththeedgeapplicationserverevenifitincludesonlyUPNFswithoutcoreCPNFs.TheadvantagesofvirtualizingallNFsongeneral-purposehardwareintheedgedatacenterareevident.BeingsituatedbetweenthecentraldatacenterandtheRANsiteallowsworkloadsfromseveralNFstobedistributedacrosseachRANsiteorconcentratedinthecentraldatacenterbasedontheavailablecomputingresourcesateachlocation.Additionally,therequiredcapacityforeachNF,includingedgeapplications,canvaryduetoseveralfactors,suchastheinitialsetupoftheedgeservice,spikesinservicesubscriptions,theadditionofnewfrequencybandsattheRANsites,orincreasedapplicationcomplexityastheedgeserviceevolves.TheflexibilityinredistributingresourceswhenadjustingtheallocationamongNFswithintheedgedatacenterissignificantlygreaterthanthatofconventionalhardware-basednetworks.Thissetupincludestheabilitytoaddmoreresourcesthroughadditionalhardwareinstallations

04

Virtualization

inDistributedRAN(D-RAN)Sites

TheNFsofaD-RANsite,composedofvariousfunctionswithintheDU,canbevirtualizedonthesamegeneral-purposehardwareplatform.Dependingonthenetworktopologydesign,theNFscorrespondingtotheCUmayalsobeincluded.Additionally,avirtualizedNFcanalsoreplacetheCSRusedtoconnecttheD-RANsitetothebackhaulnetwork.ThisstructureallowstheD-RANsitetobeconfiguredusingonlyCOTShardware.Inthefuture,asgeneral-purposeCPUscontinuetoimproveandsoftwareoptimizationadvancesaccordingly,itisanticipatedthat,inmanycases,asingleCOTSserverwillbeabletosupportbothRANfunctionsandCSRfunctionsforallofthefrequencycarriersinaD-RANsite.Furthermore,theremainingcomputingresourcescanenabletheintegrationofAIapplicationswithinRANs,eliminatingtheneedforseparateequipment.

05

BenefitsofVirtualizationin

PrivateNetworkDeployment

BenefitsofvirtualizingtheRAN,core,andtransportelementsinaCOTSservercanbeseenintheestablishmentofprivatenetworksinindustrialfactories,schools,andenterprisecampuses.Unlikenationalmacronetworks,theseprivatenetworkstypicallyconsistofdozenstohundredsofcells.Usingtraditionalhardware-basednetworkequipmentwithfixedconfigurationsandtrafficcapacitywouldrequiretheseparatemanufacturingofdedicatedequipmentforeachNFbasedontheneededcapacity.Thisapproachoftenresultsineitherunnecessarylarge-scaleequipmentoradditionalexpenses.Incontrast,avirtualizednetworkcanoperatewithjustoneCOTSservertosupportallcomponents,includingvRAN,vCore,andvCSR,evenforextremelysmallcapacities.Additionally,tomeetspecializedservicerequirements,it’spossibletocreateacustomizedconfigurationthatincreasesthecapacityofspecificNFsincertaindomainswhilemaintainingoverallflexibility.

12

3.

IntegrationofAIandMobileNetworks

Likeotherindustries,themobilenetworkbusinesshasrecentlyutilizedAItoconductresearchinvariousfields.Theseinvestigationsaimtoenhanceoperatorconvenience,userexperience,andtypicalnetworkperformance,suchasdataspeed.

ThissectiondiscussestherelationshipbetweenAIandmobilenetworks,primarilyfocusingontwomajorquestions:

Youcancheckthevideobyclickingontheimage.

WhatadvantagesandoutcomesmightresultfromcombiningAIand

mobilenetworks?Whatpreparations

shouldbemadein

advancetosupport

variousAIapplicationsinthefuture?

3.1

WhySoftware-based

NetworksarebestsuitedforAI

ForAItechnologytobeeffectivelyutilized,threekeycomponentsarenecessary:computingresources,data,andalgorithms.

WhenapplyingAIwithinamobilenetwork—excludingalgorithmsspecifictotheAIapplication’sfunctions—securingthecomputingresourcesrequiredforAIoperationswithinthenetworksystemisessential.Additionally,significantamountsofrelevantnetworkdatamustbecollectedandprovidedtoAIapplicationsformodeltrainingandinference.

First,weexaminetheissueofsecuringnetworkdata.TofullyleveragethepotentialofAI,itisessentialtocollectvarioustypesofnetworkdatasothatAIapplicationscanlearneffectively.Intraditionalmonolithichardwareenvironments,allnetworkentitiesmustreportentity-levelnetworkdatathroughastandard-definedinterface,regardlessofconfiguration.Thesenetworkentities,whichmaytaketheformofdedicatedhardwareequipment,canbebrokendownintointernallayers,includinghardwaredevices,operatingsystems,middleware,andsoftwarefunctions.

However,theexactstructureandstatusoftheseinternallayersareprimarilyaccessibleonlytotheequipmentprovider.ThissegmentationmightcreatealimitationforMNOs,whomustrelyonthesolutionsofferedbytheequipmentprovidertogathernetworkdatanecessaryforassessingthesystem’snormalcyoridentifyingissuesatthislevel.Consequently,theMNOs’abilitytodeterminethenormalstateofthenetworksystemdependsheavilyontheinformationprocessedandselectedbytheequipmentprovider.Moreover,therearechallengesrelatedtothecollectionandmonitoringofnewnetworkdata,aswellastheabilitytoperformnewanalysesandoptimizations.

13

Conversely,insoftware-basednetworksconstructedusingCOTSgeneral-purposeserversandopensoftwareplatforms,networkentitiesarecreatedbyintegratingtechnologiesacrossvariouslayers,includingoperatingsystemsandvirtualizedorcontainerizedNFs.Sincedifferentprovidersoftensupplythehardwareandsoftwareforeachlayer,theconfigurationandstatusinformationfortheselayersisaccessibletotheMNOusingthenetworkentities.ThisaccessallowstheMNOtocollectandintegratedetailednetworkdatafromallentities,functions,andtheirrespectivelayers.PleaserefertobelowFigure3.

Figure3.Unifiednetworkdatacollectionofsoftware-basednetwork

Thedatagatheredatthislevelnotonlyenhancestheunderstandingofeachnetworkentity,whichmayincludemultiplenetworkfunctions,butalsocontributestothecreationofbigdatafortheentirenetwork.Bycomparinghistoricalandcurrentnetworkdataoranalyzingtherelationshipsbetweenvariousdatapoints,theinterconnectionsbetweendifferentnetworkentitiescanbeunderstoodmoreclearlyandingreaterdepth.

Giventheextensiveandintricateinterrelationshipsrepresentedinthisbignetworkdata,thereisanincreasingneedforAIapplicationstoassistinanalyzingtheseassociations.Fromtheperspectiveofnetworkoperations,AIcanbeemployedtopredictpotentialissuesthatmaybechallengingforMNOstodetectdirectly,aswellastoidentifyrisksassociatedwiththeinterworkingofnetworkfunctions.

AIapplicationsrequirecomplexprocessingoflargedatasetsandunderstandingthesemanticrelationshipsbetweenthem.IfMNOsarerequiredtoallocatededicatedcomputingresources,suchasseparateCPUsandGPUsexclusivelyforAI,itcouldsignificantlyincreaseCAPEXandcreatesubstantialentrybarriersfortheinitialimplementationofAItechnologyinnetworks.Insoftware-definednetworks,mobilenetworkworkloadsandAIworkloadscansharethesamehardware,simplifyingtheintroductionandexpansionofAIfunctionalities.AIworkloadscanbedistributedfromcentraldatacenterstoedgedatacentersandD-RANsites.ThedecisiononwheretoplacetheAIworkloaddependsonthefunctionalattributesoftheAIapplication,includinglatencyrequirements,thesizeoftheassociatednetworkdata,andthetargetapplicationfortheanalysisresults.

AnotherimportantconsiderationishowtoallocatetheavailablecomputingresourcesateachlocationbetweennetworkfunctionsandAIworkloads.Forinstance,acentraldatacenterwiththehighestcomputingcapacitycanhandleAImodeltrainingandmanagement,whileedgedatacentersorD-RANsitescanreceivethetrainedAImodelsandfocusonspecificAItaskstooptimizeresourceutilization.

Additionally,software-definednetworkscancentralizethemanagementofdistributedcomputingresourcesacrossalltheselocationsforbothnetworkandAIworkloads.

14

3.2AI-poweredAutomation

Operatingamobilenetworkrequiresdeepexpertiseacrossmultipledomains.Asthenetworkevolves,itisdividedintovariousservices,suchasIoTservices,latency-sensitiveservices,fixedwirelessaccess,andtraditionalvoiceanddata.Inthepast,expertsmanagedandoperatedeachdomainindividual