邊緣計算演進-Evolving Edge Computing 2024

WHITEPAPER

EvolvingEdgeComputing

Contents

1WhyEvolveEdgeComputing?

2Vision

2.1EdgeVersusCloud

2.2Why‘CloudLike’inEdgeComputing?

2.3What’schanginginIoT/EdgeComputing?

2.4ChallengestoOvercome

2.5Summary

3.6Bibliography

WHITEPAPER2

1WhyEvolveEdgeComputing?

Edgecomputingisatermthathasbeeninuseforalongtime.Throughout

theindustry,therearemanyreferencestoedgeandmanypre-conceptions

aboutwhatthatmightmean.Theterm‘edge’istypicallyusedfordevicesthatexistontheedgeofanetworkandcancoveraplethoraofusecases,rangingfromtherouterinyourhouse,asmartvideocamerasurveyingaparkinglot,toacontrolsystemmanagingarobotonaproductionlineinasmartfactory.Itishardlysurprisingthenthat‘edge’isaconfusingtermwithsomanyuse

caseexamplestochoosefrom.

So,whatishappeningthatmeansthatArmiscallingforanevolutioninedgecomputing?Thispaperexaminestheconvergenceofseveralmarkettrends

thatpresentnewchallengesandopportunitiesinthisspaceandrequireustorethinkthewayforward.

Firstly,edgedevicesarebecomingconnectedtocloudservicessuchthattheyaregenerallylocatedclosetothesourceofdata.Inturn,theygenerateinsightthatfeedsnewdigitaltransformationservicesthatarehostedinthecloud.

Inthiscontext,wedefine‘thecloud’asbeingacentrallylocatedcomputeresource,typicallydatacenterbased,runninghigh-levelbusinessservices.

Theseservicesconsumeinsight(data)fromavastnumberofremotely

locatededgedevices.Asthiscloud-connectedtrendaccelerates,weseea

deepeningofthe‘relationship’betweencloudandedgedevices,suchthat

thecentrallylocatedservicesconsumingthedatahaveanever-increasing

amountofcontrolovertheedgedeviceswiththeaimofdrivingeverhigh

levelsofefficiencyinhowthesenetworksaredeployed.Althoughtheedgeisdistinctlydifferenttocloudcomputeresources,weexpecttoseedevelopersincreasinglybeingabletodevelopapplicationsatahighlevelthatare‘pushedout’totheedge,enablingdatainsightstoberefinedandtunedforvery

specificusecases.

WHITEPAPER3

Forthepurposesofthispaper,wefocuson‘frictionlessdevelopment’

asatermthatembraceshigh-levelworkloadswithhardwareabstraction,whileallowingthedevelopertoexploitthefullbenefitsoftheunderlyinghardware.

EvolvingEdgeComputing-EssentialIngredients

Developersneedtofocusonvalueadd,embracestandardsandmaximizere-use

‘Cloud-like’

Agileinnovationwithrapid

re-useacrossdevices.

Securityatscale

Trusteddevicesandtrusted

SWwithsecurelifecycleand

regulatorycompliance.

ModularSW

Complexmulti-vendorSWstacksthatworktocommonbestprectices.

Heterogeneity

Hardwareefficiencytuned

tospecificusecases.

Collaborative

Newmodelsof

collaborationtounlockthepotentialofedgecompute.

Eliminateneedlessfragmentation

Rightbalanceof

standardsandinnovation.

Eliminateunnecessarynon-differentiating

perplatformoverheadson-Arm.

Eachpartofthevaluechainfocuseson

value-addanddifferentiation.

FIG.1

EvolvingEdgeComputing–EssentialIngredients

Secondly,weseeahugeshiftinthemarkettodrivinginsightthrough

artificialintelligence.Typically,thismeanspushingAImodelsouttoedgedevicessotheycandelivertheinsightneededforbusiness-levelservices.

Finally,thesedevicesneedtobemanagedinasecureway.Asdescribedlaterinthepaper,emergingregulationsmandatesoftwaresecurityand

guaranteedupdates,makingitincreasinglyimportanttoconsiderthefullsecuritymodelofedgecomputing.Whendeployedatscale,edgedevicesareperformingacriticalroleinthedeliveryofhigh-valueservicesand

makingthemmorevulnerabletobadactormanipulation.

WHITEPAPER4

Secureidentityandsecurelifecyclemanagementarecriticalconsiderationsforabest-practiceedgecomputingapproach.

Inthecontextofthispaper,edgecomputingandsubsequently,edgeAI,

typicallyencompassescompute-richdevicesthatcanbeprogrammedin

high-levelabstractedlanguagesthatmakethemaccessibletoabroadrangeofdevelopers.FromanArmarchitectureperspective,thiscurrentlyrelies

onArmCortex-Aastheprincipalprocessingelement.Theabilitytosupportcompute-intensiveworkloadsandrichoperatingsystems,includingLinux,allowsproductsbasedonCortex-Abasedtoaddressthewidestpossible

setofusecases.

WecanexpectmanyedgeAIusecasestobepower-consumptionandcostsensitive,sothereisanongoingneedtobalancetheseaspectsacrosstheecosystem.Withthisinmind,wealsolookattheneedforheterogeneity,

i.e.,movingcompute-intenseworkloadstospecialisttypesofcomputethatofferamorebalancedapproach.

2Vision

Asuse-casecomplexityandthescaleofsmartconnectededgedevices

deploymentgrows,almostexponentially,sometechnologiesusedin

cloud-native

[1]

solutionsarebeingembracedinedgecomputing.Weseeafuturethatempowersthenextgenerationofapplicationdeveloperswithfrictionless‘cloud-like’developmentflowsthatfuelcollaboration,maximizere-use,acceleratetimetomarket,andreducethetotalcostofownership

onArm.TherapidadvancementofAIusecasesisexpectedtofuelmostofthegrowthintheedge(oredgeAI)market,withinferencebeingdeployedatscaleacrossmultiplearchitectures.

WHITEPAPER5

Thisrapidshiftinedgecomputerepresentsseveralchallenges,whichArmbelievesnecessitateanevolved,best-practiceapproachtoedgecomputingtoenabletheintelligentedgethrough:

—Re-useofsoftwarecomponents:Applicationsareakeydifferentiator.Theavailabilityandre-useofthecoreunderlyingstackiscriticalas

developerswishtofocusondifferentiationandmaximizere-useelsewhere.

—Embracingheterogeneitythroughabstractionofthecomplexityofdifferentiatedhardwarewithacommonsoftwareecosystem:

Devicesareuse-caseoptimizedbasedoncost,power,andperformance,drivinghybriddevicearchitectures(CPU/GPU/NPU/ISP,andsoon).

Thecommonsoftwareecosystemneedstoprovideanintegratedviewofthesystemwithlevelsofabstractionthatreducecomplexity.

—Genericabstracteddevelopmentflowsthatfuelcollaboration,speedtimetomarket,lowertotalcostofownershipandmaximizere-use:

Usecloud-nativederivedmethodologies,suchascontinuousintegration/continuousdeployment(CI/CD),todevelop,testapplications,anddeployefficientlytotargethardware.Developmentflowefficiencyiskeyinboththedevelopmentphase,aswellasinlong-tailmaintenanceoncethe

applicationisdeployed.

—Securityatscale:Thisisachievedthroughfrictionlesssecurelifecyclemanagementandregulatorycompliancetoreducetotalcostof

ownershipforthedeployedlifetimeofthedevice.

2.1EdgeVersusCloud

Beyondhardwareconstraints,thereareseveralkeydifferencesbetween

edge[

2

]andcloudasoperationalenvironments.Edgenodesanddevicesarepurpose-builtwithdifferentcostconstraints,resultinginmanydifferentconfigurationsdeployedovermultiplegenerationsofunderlyinghardwarecomponents.

WHITEPAPER6

Nodesdifferinhardwareresources,suchasCPUarchitecture,

micro-architecture,corecount,memory,storage,connectivity(latencyandbandwidth),peripherals,andaccelerators.Additionally,edgenodes

andgatewaysaremorelikelytorequiredynamicfrequencyscaling(eitherbecauseofbatteryconservationorthermalthrottling).Thishighdegreeofhardwareheterogeneityhasimplicationsondeployment,wheremultipleversionsofanapplicationmayberequiredtosupportdevicedifferences.

CloudNativeCloudEdge/IoTEmbedded

Highperformancecloudnativecompute

Optimisedcompute

High-performance,portableworkflowsUse-caseoptimizedefficiency,targetedworkflows

Deploy,

maintain

and

enhance

Deploy,

maintain

and

enhance

Deployandmaintaine.g.SW

updates

Deployandforget

Deploy,

maintain

and

enhance

Cloud-nativeworkflowscales

downtoedgeserver,hardwareabstractedandportable,butstill‘inthecloud.’

Embeddedsystemsscale-up,becomingsecure,connected,supportingsoftware

updatesandtakingonmoreofacloud-typedevelopmentflow.

FIG.2Organicgrowthandphysicalconstraints,suchaslocationanddifficult

CloudtransitiontoEdgeorcostlyreplacement,requiremultiplegenerationsofnodestocoexist,

leadingtodifferentSKUsofthedevicesupportedwiththesameapplicationsoftwareduringthesystem’slifetime.

Theedgeislikelytohaveahigherdatastorageandtransmissioncostcomparedtothedatacenter.Fewedgedevicesarelikelytohave

WHITEPAPER7

high-bandwidthnetworkconnections,constantconnectivityisnot

necessarilyagiven,andtransferringdatatoandfromthousandsofedgegatewaysisexpensive.Virtualmachineandcontainerimagesmagnify

thedatamovementcost,amountingtoclosetoacompletedistributiondownloadperapplication,duetoexistingpackaging.

Whilelayeredcontainerimagesareintendedtoreducethisoverhead,

third-partyapplicationpackagingmakesunderlyinglayerre-useunlikely.

Forexample,Armdevelopedaprototypehealthcareapplicationwith

machinelearning,whichused17Dockerimages,occupyingabout2.3GBofstorage.Deployingthisapplicationtothousandsofnodesovermeteredcellularnetworkingwouldnothavebeenpractical.Forthisreason,aswellasthesomewhatmoreconstrainedcomputecapability,wedonotseea

pure‘cloud-native’deploymenttoedgecomputingdevices,butrathera

frictionless‘cloud-like’modelwhichisaimedatdeliveringcloudbenefits,suchasportabilityandabstraction,inamorehardware-constrained

environment.

2.2Why‘CloudLike’inEdgeComputing?

FIG.3

BenefitsofCloudNative

Theefficienciesresultingfromminimizingtheoperationalburdenof

developers,administrators,andusersincloudcomputinghaveledtoothersegmentsevaluatingtheuseoftechnologiesoriginatingfromthecloudinotherenvironments.

WHITEPAPER8

Thedriverbehindthismovementisbasedonthelawofeconomics,namelythatthecloud-nativemodelofabstractionhasbeenshowntoaccelerate

timetomarketandsavecosts.Continuousdevelopment[

1

]isamajorcomponentofachievingafastertimetomarket.Theseadvantagesarerootedinseveralcorepropertiesofcloud-nativetechnologies:

—Portable,hardwareabstracted.

—Consistencyacrossanyinstallation/deployment.

—Timelyupdateswithoutcomplexre-integrationoverheads.

—Speedtimetomarketandmaximizere-use.—Fastapplicationdevelopmenttimes.

—Removeunnecessaryindustryfragmentationtoeliminatesiloedperplatformcosts.

2.3What’sChanginginEdgeComputing?

Digitaltransformationacrossindustriescontinuesatpace,bringingwithitnewinnovativebusinessservicesandnever-beforerealizedefficiencies.

FrombuildingthenextwaveofGigaFactoriestolow-carbon,energy-efficientcities,andtheelectrificationoftransport,acommonthemeunderliesitall—datainsightatascalenever-beforerealized.

Traditionalviewsofdatainsightarebuiltaroundadatacenter‘cloudcentric’model.Inthisscenario,sensordataissharedwiththecloud,inturnderivinginsightatscalethroughtechniquessuchasAI,todeliverthedesired

businessandefficiencyoutcomes.Thechallengecomeswithscaleandthesheernumberofconnecteddevices,andcorrespondingcomputedrives

theneedtoputprocessingclosetothesourceofthedata.Factorssuchaslatency,powerconsumption,cost,privacy,andconnectivity,alldrivethe

needtodeliverever-moresophisticatededgecomputing,ratherthansimplypushingdatatoremotecloud-basedserver.

WHITEPAPER9

Aswellasfrictionlesscomputewhereitisneeded,otherfactorsare

requiredtomeetthescaleanddemandofedgeAIgrowthoverthenextfewdecades.

Scalingdatainsightandvalue:Simplyconnectingdevicestothe

cloudbringsneitherscale,noroperationalefficiency.Traditionalcloud

datacentersdelivergenericcomputeforusebybusiness-levelapplications.Conversely,edgedevicesformthe‘real-worldinterface’anddelivermassiveinsightatscaleintothosecloud-basedservicesplatforms.Howinsight

isenabledattheedgeandhowtheseconnecteddevicesaresecurelymanagedbecomesacriticalsuccessfactorinscalingnewapplicationsandservices.

Securityatscale:Thereisgrowingregulationaroundthemanagementofelectronicdataandproducts.TheEuropeanCyberResilienceAct,

theUKProductSecurityandTelecommunicationsInfrastructureAct

andtheEuropeanRenewableEnergyDirectiveareprimeexamples.

WithsimilarlegislationprogressingintheUS,theregulatorylandscapecouldposeariskoffinancialpenaltiesandlostreputationforthosewhofailtomanagethesecurityofdigitalhardwareandsoftwareadequatelyacrossdevicelifecycles.Trustthereforebecomesasignificantfactorin

enablingscale.Edgedevicesdonotbenefitfrombeinginatraditionaldatacentersettingandareinstalledwherevertheyareneeded.

Unliketraditionalenterprisedatacentermodelswhereserversaredeployedinsecurelocationswithhighlymanagedsecurity,inedgedeployments,

weseeverydifferentdeploymentandthreatmodels.Edgedevicesmust

bedeployedinawidevarietyoflocations,withhighlyvariablesecurity

threats,e.g.,publiclylocated,susceptibletophysicalattack,connectingviapublicnetworks,tonamejustafew.Establishingtherightlevelofsecurityandtrustforedgedevicesiscriticaltoscaleapplicationsandrealizethe

businessbenefits.

WHITEPAPER10

Operationalefficiency:Aswescaleoutedgecompute,operational

efficiencybecomesakeyconsiderationwhenconsideringtotalcostof

ownership.Wecanthinkaboutthisintwoways:Firstly,thedevelopmentcosttocreatetheapplicationorservice,andsecondly,theoperationalorrunningcostsoncetheserviceisdeployed.Sinceedgecomputedevicestypicallyhavealonglifetime(5to10years,orlonger)thetotalcostof

ownershipbecomesacriticalconsideration.Thecostsincurredtooperateadeviceincludefactorssuchaspowerconsumption(linkedtorunning

costsandcarbonefficiency),aswellasdevicemaintenancecosts

relatedtomanagingsoftwareupdatesandoverallproductlifecycle.Asthedeploymentofdevicesscalesandusecasecomplexitygrows,devicevendorsandserviceprovidersincreasinglylooktooptimize

operationalefficiency.

Agileinnovation:Ourtraditionalviewofcloudcomputeisbuiltaroundagiledevelopment.Thisdeliverstremendousefficiencybothinterms

ofcloudaccessibilitytoavastnumberofdevelopersviaconsistentand

hardwareabstracteddevelopmentflows,andanagilemindsetinproductdevelopment.Asusecasesbecomemorecomplex,developersare

lookingtoembracethebenefitsof‘cloud-like’innovationinedgeusecases.Examplesincludeabstractinghardwaredifferencesasmuchaspossible

andsupportinganagiledevelopmentflowthatfacilitatesrapidinnovation,fastvirtualprototypingandcontinuousdevelopmentandimprovement

(CI/CDflows).

2.4ChallengestoOvercome

Aswehaveseen,thedemandforedgecomputeisrelentless,butsotoo

istheneedforefficiencyatalllevelsifwearetorealizethevisionatscale.TraditionalIoT-connecteddevicesthatweseetodaygosomewaytosolvingthesechallenges,butastepchangeinhowedgedevicesareenabledmust

WHITEPAPER11

happenacrossallindustries.Wecansummarizethekeychallengesasfollows:

Developa‘cloud-like’mindsetattheedge:Thetraditionaldatacenter

modelof‘writeonceandrunanywhere’doesnotmapdirectlytoedge

devicesforpracticalreasons,howeverelementsofthatmodelarecriticalforaneffectiveedgecomputingevolution.Edgedevicestendtobe

applicationspecific(e.g.asmartcamera)butmustembraceelements

offrictionlessdevelopmentforspecificbenefits.Aswethinkaboutedgecomputingasanextensionofthedatacenter,weneedawholenew

mindsetintermsofhowaccessibletheseedgedevicesaretodevelopers,andhowtheysupportagiledevelopment,virtualprototyping,and

continuousimprovements.Todeliverthisvisionalsorequiresasignificantmindsetshiftfortraditionalembeddeddevelopers.Goneisthetraditional

‘linear’developmentflowofspecifying,implementing,testing,and

deployingapplications.Instead,weshifttoCI/CD/deliveryflowtospeed

timetomarket,maximizesoftwarere-useandultimatelyreducecost.

Todothis,themarketmustbuildcommonabstractedprogrammingmodelstoopentheaccessibilityofedgedevicestodevelopersacrossplatforms,

abstractingcomplexityandlimitinghardwaredependenciesexclusivelytowheretheseaddvalue,suchasforperformanceandpoweroptimization.

Securityandprivacyatscale:Abedrockofscalingthecloudouttothe

edgeisensuringrobustsecurityandprivacy.Buildingdevicesthathave

atrustedandconsistentapproachtosecurityiscriticalfortheirlifecycle

managementandensuringtrustaroundthedevice,connection,software

lifecycle,data,andservices.Withsoftwarestacksbecomingincreasingly

complexandmultivendor,weseegreateraneedforcomposablesoftware,wherebyeachpartyownsonlytheportionofsoftwarethattheycareabout.Withinthismodel,eachsoftwarecomponentessentiallyhasitsownsecurelifecycle.Underpinningthisistheneedforconsistentplatformsecurity

capabilities,suchassecureboot,secureupdates,securestorage,

WHITEPAPER12

andtrustedcrypto.Howeachofthesoftwarecomponentscanaccessthesesecureplatformservicestomanagetheirlifecycleiscritical.

Eliminateneedlessfragmentation:Needlessfragmentationholdsback

innovationandslowsthepaceofadoptionatscale.Itisthereforeessentialtoseekoutcommonalitythatremovesneedlessnon-differentiationsothesupplychaincanfocusonlyonthedifferentiationthataddsvaluetotheirbusinessandthemarket.Anobsessiveattentiontoefficiencyisneeded

bothinthedevelopmentofthedevice,aswellastheoperationalcosts.

Amodularapproachtosoftwaredeployment:Fragmentationchallenges

extendtosoftwareasweconsidertheincreasinglycomplexusecasesfor

edgedevices.Itiscommonplaceformultivendorsoftwarestackstorun

onanedgedevicewithmanythird-partycomponentsneedingtocome

togetherandinteroperate.Increasingly,end-marketdeploymentscareaboutwhatsoftwareisrunningonedgedevices.Fleetmanagers,forexample,

wanttoknowwhatoperatingsystemsaredeployed,whatsecuritypatchesarepushedout,andwheredifferentsoftwareassetsarecomingfrom.

Thedesireforchoice,coupledwithgrowingcomplexity,isdrivingtheneedformodular,interoperablesoftwarethatcanbemaintainedthroughoutitsdeployedlifetime.

Balancestandardizationanddifferentiation:Themarketmustembracestandardsandcommonalitywherenecessarytospeedtimetomarket,

reducetotalcostofownership,andeliminateneedlessfragmentation.

CollaboratingonArmcanbringtherightlevelofstandardization,while

allowinghardwareinnovationanddifferentiationtothrive.Thereisno

single‘recipe’foredgedevicesfromanArmplatformpointofview.

Instead,weconsider‘thesetofhardwareandsoftwareinterfacesneededtominimizethecostofbooting,running,andmaintainingoperatingsystemsandothersystemsoftwarethroughthelifetimeofthedevice’.

WHITEPAPER13

Benefitsofthisapproachinclude:

—Reducestime,cost,andeffortfromgettingsoftwaretoinstallandworkfordevicelifetimes.

—Removesnon-differentiatingcostfromtheecosystem.

—Allowstheecosystemtoinvestmoretimeandmoneyonworkthataddsvalue.

Today,initiativeslike

PARSEC

forstandardizedhardware-abstractedsecurityservicesarebecomingessential,asisaconsistentapproachtosecurity,whichisprovidedby

PSACertified

.Plus,through

ArmSystemReady,welookathowoperatingsystemsaresupportedonedgedevicesasacriticalfactor,alongsidetheneedtoofferandmaintainoperatingsystemdistributionsondevicesfortheircompletelifecycle,

whileeliminatingper-platformportingcosts.

HeterogeneityinedgeAI:Whenthinkingaboutcloudnative,

weimaginecontainerizedcomputeworkloadsthatcanruninafullyportablemannerinclouddatacenters.Asweestablishedearlyinthis

document,edgecomputingtendstobeapplicationspecificandoptimizedforcertainworkloadsandpower/performancebudgets.Overthelast

fewyears,weareseeingadeepeningtrendfor‘a(chǎn)cceleratedcompute,’wherebyhardwareaccelerationisappliedtocommonandcompute-intensiveworkloads.Acceleratedcomputetakesmanyformsbut

principallyfallsintotwoareas:

01In-lineaccelerationthatoccursaspartoftheCPUoperation(e.g.,ArmScalableMatrixExtensions).

02Offloadacceleration(e.g.hardwarethatsitsalongsidetheCPU,

suchasanNPU,bprovidingheterogeneityintheprogrammingmodel).

WHITEPAPER14

Acceleratedcomputeisusedtoimproveperformance,reducepower

consumptionforspecificworkloads,orsometimesboth.Examininghow

developerexperiencesscaleacrossheterogeneousplatformsisessentialtoavoidneedlessfragmentationandsiloeddevelopmentsbecoming

deeplyentwinedtospecifichardwarevariants.Aswelooktowardsthe

evolutionofedgedevicesasoutlinedinthispaper,thepartialdecouplingofhardwareandapplicationasatrendmovesustowardan‘a(chǎn)pp-like’

modelthatfa