Table

of

contentsDriving

forces

of

6G

native

AI1236G

native

AI

architecture

designand

key

features6G

and

Foundation

Model2Demand

for

Ubiquitous

IntelligenceAI

has

become

the

core

driving

force

for

a

new

round

of

industrial

transformation.

Theautomation,

digitalization

and

intelligence

of

the

industry

require

ubiquitous

intelligence.NetworkautonomyneedsAICustomersneedAIBusinessneedsAIOperation

&

maintenance

Emergency

communicationMedical

RecognitionSecurity

monitorVoiceprint

recognitionMachine

translationSmart

NavigationPersonalized

Recommend.Robot

rescueSmart

ManufacturingSmart

coverageCustomized

networkThe

integration

of

6G

and

AI

includes

two

aspects:

"AI

for

Networks"and

“Networks

for

AI"3The

Driving

Force

of

AI

for

NetworksMobile

communication

technology

faces

bottlenecks,

requiring

urgent

technological

innovation

andinterdisciplinary

integration.

AI

is

a

key

solution

for

enhancing

network

performance.Traditional

communication

systems

faceperformance

bottlenecks6G

poses

more

challenging

demandmetricsairinterfaceMore

accurate

channelinformationMore

precise

positioningEnhanced

interferencecancellation

capabilitiesEnhanced

energy

efficiency,spectral

capacityConflicts

arise

between

network

operationefficiency,

complexity,

and

costnetworkImproved

scene

adaptationspeedCurrent

technology

falls

short

ofmeeting

6G's

needsnetwork

operation

and

maintenance

efficiencyDifficulty

in

estimating

larger

scaleMIMO

channelsMore

balanced

trafficschedulingDense

base

station

deployment

leadsto

increased

interferenceFaster

networkinginterference

avoidanceMore

complex

system

designs

lead

toincreased

energy

consumptioncontradictorytriangleMore

refined

businessidentificationComplex

routing

in

heterogeneousequipment

networksnetworkcostMore

accurate

fault

locationDiverse

communication

scenariorequirements

are

fragmentedcomplexity4The

Driving

Force

of

Networks

for

AIITU

extends

6G

scenarios

to

ubiquitous

intelligence.

AI

needs

to

be

transformed

into

new

capabilitiesand

services

for

6G

communication

networks

to

achieve

AIaaS6G

network

inherently

providesITU

extends

6G

scenarios

to

ubiquitous

intelligenceAI

services5GGet

AI

anytime,

anywherecommunicationservicecommunicationabilityLow

latency

AIinference/trainingSupport

mobile

AI6Gcommunication

computingAI

service

qualityassurance+abilityabilityAIserviceperceptionability

dataabilityAImodelabilityAI

security

and

privacy

protection5Challenges

in

the

Integration

of

5G

Networks

and

AIFulfilling

6G

and

AI

integration

demands,

the

universality

and

efficiency

of

existing

AI

design

methodsdriven

by

scenario

use

cases,

plugins,

or

grafts

need

to

be

improved.Scenario-driven

AIExternal

or

grafting

AInDesign

separate

AI

models

for

specific

air

interface

and

networkoptimization

use

casesnAdd

AI

servers

or

AI-related

network

functions

to

the

network,

such

as

NWDAFNWDAFIntelligentdataAIserversCNRANUEMassivetrafficdataanalysisAI

forNetworksChangingchannelconditionsAntennaweighttuningNetworkManagementReducedswitchingperformanceUsermovementpredictionProblem：It‘s

challenging

to

guarantee

real-time,

effective,

andconsistent

data.

Completing

the

entire

AI

process

involves

high

trial

anderror

costs.Problem:

AI

models

have

low

generalization,

longdevelopment

cycles,

and

high

costsnDesign

different

AI

service

processes

for

different

third-party

AIscenariosnCloud

AI

service

providers

provide

best-effort

AI

services

after

users

submit

ordersHigh-speedintelligentfollowingSubmitAINetworkInternetofVehiclesServiceOrderTransmissionNetworks

forAIReal-timemulti-agentsmartfactorycollaborationUsermovementNetworkUEXR/VRCloudAIServiceProviderpredictionProblem：Data

is

only

uploaded

to

the

cloud,

making

it

difficult

toefficiently

leverage

the

ubiquitous

resources

within

the

network,

whichcannot

guarantee

the

quality

and

security

of

AI

servicesProblem:

The

network

struggles

to

rapidly

deploy

AIservices

for

diverse

scenarios66G

Native

AI

Design

PrinciplesTo

achieve

ubiquitous

intelligence,

6G

network

architecture

requires

"four

transformations"CloudCloudCloudAIprovidersNWDAFCNCNCommunicationQoSAI

workflow

15G

External

AI6G

Native

AI

AI

workflow

2CommunicationQoSTrafficanalysisTrafficanalysisAntennaadjustmentAntennaadjustmentMovementpredictionMovementpredictionCommunicationComputingAIFour

elements

collaborationDataAlgorithm7Table

of

contentsDriving

forces

of

6G

native

AI1236G

native

AI

architecture

designand

key

features6G

and

Foundation

Model86G

Native

AI

network

architectureChallenge:

As

the

three

fundamental

components

of

AI

(data,

algorithms

and

computing)

have

gainedsignificance

on

par

with

network

connections,

the

design

of

the

corresponding

architecture,

interfaces,

andprotocols

should

span

the

entire

AI

lifecycle.Dataplane:managesnetworkdata

andprovidesdata

servicesResourcelayer:provideunderlyingresourcesComputingplane:managescomputingandprovidescomputingservicesNetworkfunctionlayer:providespecificnetworkfunction/networkservicecapabilitiesIntelligentplane:providestheoperatingenvironmentfor

fulllife-cycleofnativeAI.Applicationandservicelayer:providecorrespondingsupportforcustomers'businessneeds.Unlike5Gnetwork,newdata

plane,smartplane,andcomputingplanewillbedefinedin6Gnetwork,andtraditionalcontrol

planeanduserplaneare

expectedto

beextendedaswell.96G

Native

AI

network

architectureTo

achieve

AI

services,

collaboration

among

communication,

data,

computing,

and

intelligence

is

essential.

Theintegrated

architecture

enables

east-west

and

north-south

processes,

meeting

QoAIS

requirements

for

internal

andexternal

AI

services1.

Ubiquitous

intelligent

network

and

AIintegrated

service

framework2.

From

External

Overlay

to

InternalSynergy3.

From

best-effort

to

on-demandNetworks

for

AIIntelligence

PlaneCommunicationPlaneData

PlaneCompute

PlaneDigitalTwinAIManagementandOrchestrationAI

Use

CaseGenerationand

StrategyNetworkfunctionarrangementAI

forNetworksService

QoSAnalysisAITaskManagementDigitaltwinforcommuni-cationFourAI

TaskLifecycleManagementAI

TaskSchedulingElementResourceControlCommunicationControlComputingControlDataControlModelControlCommunicationBearerDataProcessingComputingExecutionModelOperationResultsofAItraining,inference,compression,verification,etc.10Key

Feature

1:

AI

Service

Quality

(QoAIS)Traditional

QoS

systems

primarily

emphasize

session

and

connection

performance,

lacking

comprehensive

supportfor

diverse

requirements；The

QoAIS

indicator

system

incorporates

security,

privacy,

autonomy,

and

resourceoverhead

as

new

evaluation

dimensions

to

form

a

standardized

AI

service

quality

evaluation

system.QoAISGuaranteeMechanismSmartEntertainmentSmart

CitySmartIndustrySmart

LifeSmartCommunityPlatformizedService

NetworkManagement&AIServiceService

QoSOrchestrationAITaskTask

QoSTaskManagementAlgorithmComputinDataConnectionResourceQoSTask

ControlgUnified

IP

computing-network

baseOTN/OXCOTN/OXCOTN/OXCAll

optical

baseComputing-Network

InfrastructureKey

Feature

2:

Deep

integration

of

AI

computing

and

communicationDesigning

a

native

AI

protocol

that

integrates

computing

and

communication

is

necessary

to

meet

AI‘s

connectivityand

distributed

computing

service

needs.It

is

achieved

through

three

dimensions:

Management

Plane,

Control

Plane

and

User

PlaneControl

Plane:

Three

Modes

of

Deep

Convergence

ofComputing

and

CommunicationManagement

PlaneMode1Mode2Mode3FunctionalarrangementQoSanalysisComputing

requirementsfor

6G

native

AICoordinationxNBxNBHigh

computationalefficiencyConnectioncontrolComputingcontrolConnectioncontrolComputingcontrolConvergedcontrolLow

energyconsumption

andlatencyCCBCEBCCBCCBCEBCEBComputing

Task

Data

Transmission&ExecutionMeet

the

differentiatedQoAIS

needsTask1CEBCEBCEBUser

PlaneCCBCCBTask3collaborativedesignofcomputingandcommunicationprotocolCCBCCBCEBCSCEBTask2CEBCEB：ComputingExecutionBearerCCB：ComputingConnectionBearerCS：ComputingSession=CEB+CCBKey

Feature

3:

Data

Generation

and

Reliable

AIThe

massive

training

data

demand

and

high

risk

of

trial

and

error

for

AI

in

the

network

require

networkdigital

twins

to

achieve

on-demand

data

generation

and

reliable

AI

and

verificationData

generation

andNetwork

Digital

Twinoptimization1.

Reduce

the

cost

of

data

collectionandtransmission;2.Solveproblemssuchasdifficultyinobtainingtraditionalrealdata;3.Technology:

DataAugmentationinGANs;Network

virtualsceneNetworkstatepredictionNetwork

AIRequirementsAIservicesDigital

twinentityDigitaltwinmodelingrequirementsPre

validation

of

AIExternal

demandAuto-generatedrequirementsProcesseddataData

on-demandcollection

andgeneration1.Intendedto

completeperformancepre

validationwithoutaffectingnetworkoperations;Requirementsfordatacollectionandgeneration2.

Reduce

potential

risks

thatdecisions

maylead

to,

such

asdeterioratingnetworkperformance;RadioCUDUAAUphysicalnetworkVirtualizationCore

NetworkCloud

based

Radio

Access

NetworkTable

of

contentsDriving

forces

of

6G

native

AI1236G

native

AI

architecture

designand

key

features6G

and

Foundation

Model14The

Convergence

of

6G

and

AI:

A

New

Era

of

Foundation

ModelsAs

AI

enters

the

era

of

general

intelligence,

the

emergence

of

Foundation

Models

promises

a

profoundtransformation

in

the

integration

of

6G

and

AINetworks

forFoundation

ModelsFoundation

Modelsfor

NetworksThenetworkservesasaplatformto

supportorprovideFoundationModelsservicesFoundationModelswillenhancemobilenetworkservicesinaspectssuchasoperations,execution,andverificationDomainsRequirementsImpactonNetworksSmallNetworkOperationsMulti-modalMachineLearning,LanguageUnderstanding,TextGenerationNetworkMaintenanceNon-standardDataGovernance,DataAlignment,NaturalLanguageUnderstanding,CodeGenerationMediumLargeNetworkRunningNon-standardDataGovernance,ImageGeneration,VideoGeneration?

Providerichenvironmentaldataforfoundationmodels?

Offerintent-based

servicesto

users?

Achieveglobalcollaborativecontrolof

intelligentterminals?

DetectingFailures

andGeneratingSolutions?

OrchestratingandSchedulingTask

Workflows?

PlayingaVitalRole

intheValidation

Phase15Networks

for

Foundation

Models6G

native

AI

facilitates

the

training

of

foundation

models

by

providing

links

and

data

services

during

the

trainingprocess,

and

supports

the

inference

process

with

links,

computation,

and

model

decomposition/distribution

servicesGAI

training

servicesAI

inference

servicesProcesseddataProcesseddataMassivedatacollectionInferencerequestsDataprocessingAIinferenceUE6GNetworkCloudAIprovidersUE6GNetworkCloudAIprovidersFoundationmodelstrainingoftenneedshigh-speedfiber

Foundationmodelsrequiresignificantstoragespaceandopticconnectionsindatacenters,makingradionetwork

powerfulAIinferencechips,whichcannotbemetbyaFeaturesServicesdeploymentchallenging.singlebasestation.Collectinguserandnetworkdata,preprocessingit,andmanagingtrafficto

supportmodeltrainingWithpropermodelsegmentation,modelscanbedeployedinwirelessnetworksto

offerAIinferenceservices.PotentialgainsIn6Gnetworks,deployingmodelscloserto

userscanreducelatency6Gnetworksprocessdataefficiently,reducingdatatransmissionandimprovingcloudAItrainingformodelsHowto

balanceincreasedinferencelatencywithreducedtransmissionlatencyin6Gnetworks?Are

techniqueslike

modelsegmentation,compression,andaccelerationfeasibleforFutureissuesTherequiredspecialdataanalysistechniques?Howtoefficientlyscheduledatainadistributed?16models?databeeffectivelyscheduledbetweennodes?Foundation

Models

for

NetworksFoundation

Models

for

Networks

face

significant

challenges

due

to

the

abundance

of

structured

data

and

unclearcommonalities

among

different

network

problems,

unlike

ChatGPTExploring

in

phases,

beginning

with

the

exploration

of

network

operations

ai

general

modelsProgressing

from

small-scale

to

large-scale

and

from

offline

to

real-time,

ultimately

investigating

thefeasibility

of

unificationSmall-scalelarge-scaleunifiedOfflineScenario-basedoperationmodelOperation

universalmodel？Networkuniversal

modelsmallmodel1Service-levelrunningmodel？smallmodel2…smallmodelNMulti-scenariouniversal

runningmodelNetwork-levelrunningmodel？Single-systemrunningmodelReal

time17The

Challenges

of

Network

AI

Foundation

Models

-

DataNetwork

operation

and

maintenance

data

is

mainly

available

at

minute/hour

intervals

from

a

consistent

source,while

network

operational

data

is

more

complex

due

to

varying

time

intervals,

standardization,

and

data

sources,making

it

harder

to

acquire.Data

openness

and

standardizationIndustry-wide

collaborative

data

openness6GANA

collaborates

with

multiple

organizations,

including

the

NineHeavens

platform,

to

release

four

major

datasets,

creating

an

industry

datasharing

ecosystem

to

support

network

AI

research!Difficult

dataacquisitionPoor

dataqualityIntelligent

RANSlicing

DatasetCSI

CompressionFeedback

DatasetData

opennessContinuouslycuratingandaccumulatingintelligentnetworkdatasets,opento

thepublic,to

buildaseriesof

innovativesmart

networkecosystems,andsupportresearchstandardizationCollaboratewiththeindustryto

jointlyformulatenewdatacollectionstandardsanddevelopadynamicdatacollectiongranularityschemetailoredto

specificneedsNetwork

AI

SchedulingTechnology

Research

DatasetRadio

ResourceScheduling

Dataset18The

Challenges

of

Network

AI

Foundation

Models

-

EvaluationEstablishing

a

comprehensive

evaluation

system

for

AI-enabled

networks

to

ensure

overall

practicality,

balance,

andsystematicity

of

technical

solutions.RequirementsChallenges?

Creatingaholisticmetricsystembalancingcapability,efficiency,andquality.?

Coveringcustomization,generalization,universality,real-timeperformance,reliability,andcost-effectivenessFeasibility

ofQuantitativeMetricsMetricsDrivers?

Designinganengineeringevaluationmethodencompassingtheentirelifecycleandallelements?

Includingtraininginference,elementweights,comprehensiveobjecti