高盛-新的AI技术能否加速自动驾驶落地-2024-08-自动驾驶_第1页
高盛-新的AI技术能否加速自动驾驶落地-2024-08-自动驾驶_第2页
高盛-新的AI技术能否加速自动驾驶落地-2024-08-自动驾驶_第3页
高盛-新的AI技术能否加速自动驾驶落地-2024-08-自动驾驶_第4页
高盛-新的AI技术能否加速自动驾驶落地-2024-08-自动驾驶_第5页
已阅读5页,还剩75页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

10July2024|5:03PMEDT

Someinvestorsandmarketobserversarguethattheindustryisonthecuspofa

MarkDelaney,CFA

+1(212)357-0535|

breakthroughinautonomousvehicle(AV)scalingdrivenbynewAIandGPU

mark.delaney@GoldmanSachs&Co.LLC

technology,whileothersarguethatbroad-basedAVdeploymentsmayneveroccur.

KotaYuzawa

+81(3)4587-9863|kota.yuzawa@

Thebottom-lineisthatwebelieveimprovedAItechnologywillhelpthe

GoldmanSachsJapanCo.,Ltd.

industryreachhigherlevelsofperformance,althoughwealsobelievethat

AllenChang

+852-2978-2930|allen.k.chang@

widescaleAVadoptionisstillatleastafewyearsawayasabasecase.We

GoldmanSachs(Asia)L.L.C.

believethatgloballevel3(L3)ADASpenetration(e.g.vehiclesthatcan

GeorgeGalliers

+44(20)7552-5784|

situationallyhaveeyes-offdriving,suchasonahighway)willreach10%ofthe

george.galliers@

GoldmanSachsInternational

marketfornewvehiclesalesin2030,withlevel4orAVs(e.g.eyes-offina

TinaHou

givenarea,suchasarobotaxiinacity)at2.5%in2030.Thisimpliesthatlevel4

+86(21)2401-8694|

tina.hou@

(L4)volumeswillcontinuegrowing,albeitgenerallyforcommercialusecases

GoldmanSachs(China)SecuritiesCompanyLimited

likerobotaxisintheneartointermediateterm.

EricSheridan

+1(917)343-8683|eric.sheridan@

WeassumemostoftheADAS(advanceddriverassistancesystems)andAVindustryvolumegrowthinthenextfewyearswillcomefrompartially

GoldmanSachs&Co.LLC

ToshiyaHari

+1(646)446-1759|toshiya.hari@

autonomousL2/L2+vehiclesthatrequiredriversupervision.Weassume

GoldmanSachs&Co.LLC

DaikiTakayama

L2/L2+mixwillrisefromabout20%ofsalesthisyeartoabout30%in2027.

+81(3)4587-9870|

daiki.takayama@

GoldmanSachsJapanCo.,Ltd.

OurAVforecastimpliesthataglobalfleetofafewmillioncommercialAVs

BenMiller

usedforridesharecouldbeontheroadin2030.Althoughthiswouldcomprise

+1(917)343-8674|

ler@

lessthan1%oftheglobalcarparcofover1bnvehicles,itcouldresultina

GoldmanSachs&Co.LLC

>$25bnmarketforpersonalmobilityfromrobotaxis.

LincolnKong,CFA

+852-2978-6603|lincoln.kong@GoldmanSachs(Asia)L.L.C.

WebelievethatstocksinvestorsshouldownonthisthemeincludeNvidia,Uber,

JerryRevich,CFA

Mobileye,Renesas,Baidu,DesaySVandQuanta.

+1(212)902-4116|jerry.revich@GoldmanSachs&Co.LLC

GivenadvancesinAItechnology,includingthelatestNvidiaprocessorsaswellas

DanielaCosta

+44(20)7774-8354|

differenttrainingapproaches(e.g.afully“end-to-end”approachofcamera/sensor

daniela.costa@

GoldmanSachsInternational

inputsinanddrivingpolicyoutthatcanpotentiallyhelpsolvedifficultedgecases

VerenaJeng

especiallyifthereisenoughdata,ora“compoundapproach”thatutilizesmachine

+852-2978-1681|verena.jeng@GoldmanSachs(Asia)L.L.C.

learning/AIbutwithsubsystemsthatcanallowforefficiencyandverificationofthe

RyoHarada

solution),weattempttobetterunderstandiftherateofprogresstoward

+81(3)4587-9865|ryo.harada@GoldmanSachsJapanCo.,Ltd.

wider-scaleadoptionofautonomywillaccelerate,includingL3andL4.Seeour

AlexanderDuval

+44(20)7552-2995|

GoldmanSachsdoesandseekstodobusinesswithcompaniescoveredinitsresearchreports.Asaresult,

investorsshouldbeawarethatthefirmmayhaveaconflictofinterestthatcouldaffecttheobjectivityofthis

report.Investorsshouldconsiderthisreportasonlyasinglefactorinmakingtheirinvestmentdecision.ForRegACcertificationandotherimportantdisclosures,seetheDisclosureAppendix,orgoto

..i//eS.html.Analystsemployedbynon-USaffiliatesarenotregistered/qualifiedasresearch

GlobalAutos&IndustrialTech

10July20242

“AI101”sectionofthisreportformoredetailsonAIapproacheslikeend-to-end.

WiththisreportweupdateourglobalADASandAVforecastsbasedoninputsfromourglobalauto,industrial,andTMTteammembers.

Keyworkinthisreport-

1)ExaminationofTesla’stechnicalprogress,andprogressofvariousAVcompanyefforts

2)IllustrativecostpermileforanAVrobotaxibusinesswithaverticallyintegratedmodel,andadiscussiononrideshareandAVbusinessmodels

3)UpdatedglobalADAS/AVforecast

4)Discussionofnewelectronicarchitecturesandimplicationsforsemis

5)2030EPSscenarioanalysisforTesla

PMSummary

Wethinkitisnotablethattherearenowautonomousvehiclesontheroad.ThereareasmallnumberofAVsoperatinginpartsofmajorcitiessuchasSanFrancisco,

Phoenix,BeijingandWuhan.

However,thetechnologyhasyettobebroadlydeployed.Keyissuesgatinggrowtharerelatedtounderstandingcomplextraffic/drivingscenarios(oredgecases),along

withbusinessandregulatoryfactors.WhileAVsfromcompaniessuchasBaidu(ApolloGo),WaymoandPony.aimayalreadybesaferintermsofaccidentspermilethanhumandriverswithingeofencedareas(e.g.perdatafromleadingcompaniessuchasWaymo,albeitwithAVsontheroadtodaytypicallyabletoqueryahumaninaremotelocationforassistanceifneeded),therecontinuetobecaseswhereAVsgetconfusedorstuckinscenariosahumancouldlikelynavigate.

Weseektobetterunderstandwherethetechnologyandindustrycurrentlystands,andifnewAItechnologycanhelptoaccelerateprogresstowardwider-scaleadoptionofL3(e.g.situationallyeyes-offandhands-offdriving,suchasonahighway)orL4(e.g.

eyes-offinagivenarea,suchasarobotaxiinacity)autonomy.

ResearchonAIscalingdoessuggestthataddedcompute,largertrainingdatasets,

andimprovedmodelarchitecturesshouldcontributetobetterAImodel

performance.MicrosoftCEOSatyaNadelladuringakeynoteatMicrosoftBuild2024notedthatsimilartohowMoore’sLawdrovetechnologyinthepast,AItechnologynowallowscomputeperformancefortrainingdeepneuralnetworkstoroughlydoubleevery6months(referencingresearchfromEpoch).

WeconsiderTeslatobeoneoftheleadersinautonomoustechnology.Tesla’ssupervisedfullselfdriving(FSD)technology,whichisanL2/L2+systemasit

requiresthedrivertopayattentionandbepreparedtotakeoveratalltimes,isalreadysaferthanvehiclesdrivenmanuallyintermsofaccidentspermiledatafromTesla(Exhibit1).Tesla’sdatashowsthatin2023accidentswithsupervisedFSD

GlobalAutos&IndustrialTech

10July20243

(whichisTesla’ssolutionthatwillallowthecartodopointtopointnavigation,andworksonhighwaysandcitystreets)occurredevery4to5millionmiles,andaccidentson

Autopilot(whichincludesmorebasicfeaturesliketrafficawarecruisecontrolandlanekeeping)occurredevery5to6mnmiles(althoughthesearegenerallyhighwaymiles),comparedtoonceevery600-700KmilesfortheUSonaverage.In1Q24,accidentswithAutopilothappenedonceevery7-8millionmiles.

Exhibit1:TeslaADASsystemmilesdrivenbeforeanaccident

Milesdrivenbeforeaccident(mns)

6

5

4

3

2

1

0

201820192020202120222023

TeslaFSDBeta/SupervisedTeslaAutopilotTeslanon-AutopilotUSaverage

Source:Companydata,FHWA,NHTSA,GoldmanSachsGlobalInvestmentResearch

Whileweconsiderthesesafetystatisticstobeimportant,measuringifTesla

driversgetinfeweraccidentswhileusingthetechnologywithactivedriver

supervision(andtakingoverifthereisanissue)isnotthesameaswhetherthevehicleswouldbebetterand/orsaferwhenunsupervised(e.g.L3/L4capability).

Wethereforebelieveitisalsohelpfultogaugethepercentageofdriveswithno

interventionsandinterventionspermiletoassesswhenTeslamayreachunsupervisedcapability(e.g.L3orL4).Thereareotherperformancemetricsbeyondjustaccidentstoconsider(e.g.properlaneselection,smoothdrivingbehavior,andrespondingto

emergencyvehicles).

WhileTesladoesn’tdiscloseinterventiondataonFSD(andinterventionsareatthe

discretionofthedriverwhichcancomplicateananalysisofthissort),somedrivers

submitinterventiondatatoTeslaFSDtrackerandthisshowsthatthepercentageof

driveswithoutaninterventionhasgoneup,andabout70%ofdriveswiththelatestversionofFSDhavebeenno-interventionperthiscrowd-sourceddata(Exhibit2).Tesla’slatestFSDversion(V12)wasdevelopedwithmorerelianceonAI,oran

“end-to-end”approachofvision/camerainputsinanddrivingpolicyoutthatcanpotentiallyhelpsolvedifficultedgecasesespeciallyifthereisenoughdata.

Moreover,thissamedatasuggestscriticalinterventionsoccurevery~300miles

withFSD.Whilewethinkitwouldbewrongtoassumethateverydisengagementwouldhaveresultedinacrash,italsosuggeststhetechnologycouldbesome

timeofffrombeingL3orL4giventhe600-700KmilestraveledbetweenaccidentsforthetypicalvehicleintheUS.OurownrecentridesinFSDenabledvehiclesalso

suggestthatFSDisimpressivebutnotyetreadytobeL3orL4,inouropinion.

10July20244

Exhibit2:PercentofTesladriveswithoutaninterventionpercrowdsourceddata

100%90%80%70%60%50%40%30%20%10%

0%

Feb-22

May-22

Aug-22

Nov-22

Feb-23

May-23

Aug-23

Nov-23

Feb-24

May-24

Aug-24

Nov-24

Feb-25

May-25

Aug-25

Nov-25

Feb-26

May-26

Aug-26

Nov-26

Feb-27

May-27

Aug27

%ofdriveswithnodisengagements--------SteadyFSDscalingRapidFSDscaling

----FSDhitsceiling

Source:TeslaFSDTracker,GoldmanSachsGlobalInvestmentResearch

ExtrapolatingtherateofnointerventiondrivesimpliesinourviewthatTeslacouldreachL3onhighwaysatfullspeeds,atleastinclearweather,inthenexttwotothreeyears,especiallyifTesla’ssignificantinvestmentsintechnology(Teslaisspending$3-$4bnonNvidiacomputethisyear)anduseofanend-to-endAIapproachhelp.Giventheaddedcomplexitiesofoperatingindenseurbanenvironmentsandhighbarforsafetyin

unsuperviseddriving,weassumegeneralizedL4wouldtakelongertoreach.Webelievethatutilizinghumanassistanceinaremotelocation(similartocurrentrobotaxieffortslikeWaymoandBaidu)couldallowTeslatoreachL4functionalitysoonerthanitis

otherwisetrackingtoalbeitwithaddedcostandscalingchallenges.

LookingatAVeffortsbesidesTesla,whileaccidentdatahasbeenpromising(e.g.

Waymo’sanalysisshowsatleast57%fewervehiclecrashespermilethanahuman

driverwithitsL4robotaxis),thecurrentdeployments(e.g.fromWaymo,Pony.aiandBaiduApolloGo)havethusfarbeentargetedintermsofdeployments,andlimitedtocertainsectionsofselectcities(wehavemoredetailsinthe“CompanyADAS

andAVefforts”sectionofthisreport).Whilethismaybedueinparttotechnologyscalabilityissues(e.g.relyingondetailed3-Dmaps,needinghumanremoteassistanceattimes,anddifficultyinhightraffic/complexscenarioslikeconstruction),wealso

believeeconomicconsiderationsareafactor.AsweshowinExhibit3,thecostsper

milearelikelyveryhighatlowvolumesdueinparttothecostofthehardware/computeandhumanremoteassistance(notethatTeslacouldhaveacostadvantagewithits

AV/ADAStechnologygivenitsverticalintegrationforinferencechips,scale,anditslimitedsensorsuite).

Assumingthatdepreciationandinsurancecostsnormalizetolevelsonparwithhumandrivencommercialrideshareentities,weestimatethatvehicledrivingcostspermileforanAVat50-75kmilesdrivenperyearpervehicleandroughly10carsperremote

operatorcouldreach~$1.00permile(weshowthisforillustrativepurposesoccurringinthe2030timeframeinExhibit3).Costswithcorporateoverhead/R&Dwouldbehigher.Longer-termthesecostscoulddecline.

GlobalAutos&IndustrialTech

10July20245

Exhibit3:IllustrativecostmodelforaverticallyintegratedAVridesharecompany

2023

2024E

2025E

2030E

2035E

2040E

AVcostpervehicle($US)

125,000

100,000

85,000

50,000

50,000

50,000

Milesdrivenpercar

22,500

25,000

27,500

75,000

100,000

125,000

Vehiclesinserviceyearend

177

259

478

2,570

18,597

72,967

Wagesperremoteoperator

76,875

78,797

80,373

87,870

94,661

99,982

Carsperoperator

3

3

3

10

30

35

Vehicledrivingcosttpermile

$3.35

$3.13

$2.94

$0.98

$0.70

$0.58

R&D($USmn)

825

908

998

1,521

1,960

2,502

SG&A($USmn)

230

265

304

612

1,127

1,479

Tottalcosttpermile

$268.71

$184.11

$102.03

$12.04

$2.36

$1.02

Source:GoldmanSachsGlobalInvestmentResearch

Forcontext,Lyftcommittedtopayingdriversatleast70%ofriderpaymentsperweekafterexternalfeeslikecommercialinsurance(weestimate~$0.30permile)are

subtracted,andLyftestimatesthatthereare~$0.31/mileofexpensesassociatedwithoperatingthecarforthedriver(i.e.fuelcosts,maintenance,cleaning,anddepreciation).WenotethattheaveragecosttoownapersonalcarintheUSis~$0.80permileperAAA,assuming15Kmilesdrivenperyear.

Inaddition,ourrecentdiscussionswithAVoperatorsinChinasuggestthatAV

companiessofarneedtochargeadiscounttoconventionalrideshareplatformssuchasDiDiofatleast30%inordertoattractuserstothenetworkandtocompensatefor

factorslikelimitstothelocationsthatAVscanpresentlyreachduetogeofencingrestrictions.

Finally,regulatory/liabilityandsocietalconcernsmaygatetherateofgrowth.For

example,Cruisepausedcommercialoperationsafterasevereaccidentinvolvingoneofitsrobotaxis,andcommunitiesmayhaveamuchhighersafetybarforAVsthancurrenthumanperformance.

AsabasecasewenowassumethatgloballyL3enabledvehiclescouldreach10%ofindustryunitsalesvolumesin2030,andthatL4willbe2.5%in2030drivenbycontinuedtechnologicaladvancement(includingfromAI)andlowercostsofrelevanthardware(e.g.lidarandtheintroductionofpurposebuiltAVplatforms).However,ifthelatestAItechnologyhelpstheindustrytoaccelerateautonomousvehicledevelopmentfasterthanweexpect,thenwebelievethiscouldoccurafewyearsearlier.

GlobalAutos&IndustrialTech

10July20246

Exhibit4:GlobalADASandAVL3-5penetrationrateasapercentofnewlightvehiclesales

Globallevel3/4/5penetrationrate

70%

60%

50%

40%

30%

20%

10%

0%

BullBase

Source:Companydata,GoldmanSachsGlobalInvestmentResearch

OurAVforecastimpliesthataglobalfleetofafewmillioncommercialAVsused

forridesharecouldbeontheroadin2030.Althoughthiswouldcompriselessthan1%oftheglobalcarparcofover1bnvehicles,itcouldresultina>$25bnmarketforpersonalmobilityfromrobotaxis(dependingonfactorssuchasASPs,tripsper

day,andaveragemilestraveledpertrip).

Exhibit5:Weestimatethemarketin2030forrobotaxiscouldbe>$25bn

Revenuepertrip

Tripsper

robotaxiperday

150

GlobalAVsinoperation(000s)

1,3502,0002,650

2030marketscenariosforrobotaxis($mn)

750

3,300

4,000

$5

2

$548

$2,738

$4,928

$7,300

$9,673

$12,045

$14,600

4

$1,095

$5,475

$9,855

$14,600

$19,345

$24,090

$29,200

6

$1,643

$8,213

$14,783

$21,900

$29,018

$36,135

$43,800

8

$2,190

$10,950

$19,710

$29,200

$38,690

$48,180

$58,400

10

$2,738

$13,688

$24,638

$36,500

$48,363

$60,225

$73,000

12

$3,285

$16,425

$29,565

$43,800

$58,035

$72,270

$87,600

14

$3,833

$19,163

$34,493

$51,100

$67,708

$84,315

$102,200

$7

2

$767

$3,833

$6,899

$10,220

$13,542

$16,863

$20,440

4

$1,533

$7,665

$13,797

$20,440

$27,083

$33,726

$40,880

6

$2,300

$11,498

$20,696

$30,660

$40,625

$50,589

$61,320

8

$3,066

$15,330

$27,594

$40,880

$54,166

$67,452

$81,760

10

$3,833

$19,163

$34,493

$51,100

$67,708

$84,315

$102,200

12

$4,599

$22,995

$41,391

$61,320

$81,249

$101,178

$122,640

14

$5,366

$26,828

$48,290

$71,540

$94,791

$118,041

$143,080

$9

2

$986

$4,928

$8,870

$13,140

$17,411

$21,681

$26,280

4

$1,971

$9,855

$17,739

$26,280

$34,821

$43,362

$52,560

6

$2,957

$14,783

$26,609

$39,420

$52,232

$65,043

$78,840

8

$3,942

$19,710

$35,478

$52,560

$69,642

$86,724

$105,120

10

$4,928

$24,638

$44,348

$65,700

$87,053

$108,405

$131,400

12

$5,913

$29,565

$53,217

$78,840

$104,463

$130,086

$157,680

14

$6,899

$34,493

$62,087

$91,980

$121,874

$151,767

$183,960

Source:Companydata,GoldmanSachsGlobalInvestmentResearch

TherearealsosignificantsocietalbenefitsfromADASandAVtechnology,asabout40kpeopledieintrafficfatalitiesintheUSannuallyperNHTSA,andover1millionpeopledieeachyeargloballyinaccidentspertheWorldHealthOrganization.

GlobalAutos&IndustrialTech

10July20247

AI101-WhatisAI,andwhat’sthedifferencebetween“end-to-end”anda“compoundapproach”?

Inthissectionofthereport,weexplainkeyconceptsforartificialintelligence,asdetailedinapriorreportledbyToshiyaHari.

ArtificialIntelligence:Artificialintelligencedescribesascienceofsimulatingintelligentbehaviorincomputers.Itentailsenablingcomputerstoexhibithuman-likebehavioraltraitsincludingknowledge,reasoning,commonsense,learning,anddecisionmaking.

Machinelearning:Machinelearningisabranchofartificialintelligenceandentails

enablingcomputerstolearnfromdatawithoutbeingexplicitlyprogrammed.For

example,thecomputerlearnshowtoidentifyanobjectsuchasadogoracatwithdata.

Neuralnetworks:AneuralnetworkinthecontextofAI/machinelearningdescribesatypeofcomputerarchitecturethatsimulatesthestructureofahumanbrainontowhichAI/machinelearningprogramscanbebuilt.Itconsistsofconnectednodesinaggregatethatcansolvemorecomplexproblemsandlearn,liketheneuronsinahumanbrain,suchasinExhibit6.Theprocessofbackpropagationisusedinmachinelearningto

adjusttheweightoftheneuronsintheneuralnetandstrengthenthepathstoproduceacorrectanswer(e.g.toidentifyanobject).

Exhibit6:Illustrativeneuralnetframework

Source:GoldmanSachsGlobalInvestmentResearch

Deeplearningisasubsetofmachinelearningwithahierarchyoflayersinaneuralnet,withdeeplearninghavingmorelayers.Forexampleinthedogorcatexample,differentlayerscouldcorrespondtothekeydefiningfeaturesofagivenanimal.

CompoundAIsystemvs.end-to-endapproach:TherearedifferentapproachestoAIdevelopmentinthecontextofautonomousdriving,withonebeinganend-to-enddesign(asdescribedmorefullyinthisarticle),whichistheapproachTeslahasmovedtoandisasinglemodelthattakesthevisioninputsinanddirectlyoutputsthedrivepolicylike

steeringandbraking.Itreducestheneedforhumancoding,andmayhelpsolvethe

GlobalAutos&IndustrialTech

10July20248

kindofedgecasesthathavelimitedAVprogressthusfar.AnalternativeapproachisacompoundsolutionthatcanutilizeAIfordifferentsubsystemsorinputs.Thesolutioncanthenusethistogetherwithgluecodeand/oroverlayotherrules/factors(e.g.

ensuringcertaintrafficlawsareobeyed).WhilethereisadebateifthiswilllimitwhatAIcanachieveintermsofhardtounderstandedgecasescenarios,itcanhaveefficiencies(e.g.ChatGPTdoesn’tneedtouseAItrainingtosolvebasicmath,itcanjustquerya

calculatormodule,asdescribedinthisblogontheprosandconsofanend-to-endvs.acompoundapproach)andacompoundsystemcanbeeasiertounderstand/verify(whichisespeciallyimportantindrivingasamistakefromAIcouldbefatal).CompoundAI

approachesstillmakeuseofadvancedAItechniquesliketransformers.Wealsonote

thatablendofapproachescanbeused,suchasanend-to-endmodelasaconsiderationforthedrivingdecisionsbutwithcertainpolicyrulesthatareprogrammed.

Willmorecomputeandanend-to-endapproachhelpTesla’sFSDprogress?

Teslahadexpectedtohavefullyautonomousvehiclesreadyin2020per

commentsatits2019AutonomyInvestorDay,andtheindustrymorebroadlyhasstruggledtomeetitsAVtargets.InthissectionwediscussifnewerAItechnologycanhelpTeslatomeetthisobjective.

Specifically,inMay2023,TeslaannouncedonXthatitwaswouldadoptanend-to-endAIapproachwithVersion12ofitssupervisedFullSelfDriving(orFSD)product.Recallthatanend-to-endapproachisvisionin(fromthecameras)anddrivingpolicy(suchassteeringandbraking)out.Moreover,onits1Q24earningscall,Teslacommentedthatitwasnolongercomputeconstrained.

ResearchonAIscalingdoessuggestthataddedcompute,largertrainingdatasets,andimprovedmodelarchitecturesshouldcontributetobetterAImodelperformance.

MicrosoftCEOSatyaNadelladuringakeynoteatMicrosoftBuild2024notedthatsimilartohowMoore’sLawdrovetechnologyinthepast,AItechnologynowallowscomputeperformancefortrainingdeepneuralnetworkstoroughlydoubleevery6months

(referencingresearchfromEpoch).

Oncompute,Teslaisexpectingtohavecommissionedandinstalledaround85KH100GPUsfromNvidiabytheendof2024,upfrom35Kasofits1Q24earningscall.Teslaplanstoinvestabout$3-$4bninNvidiahardwarein2024.TheH100improves

performancebyupto9XforAItrainingandupto30XforAIinferenceovertheprior

A100GPUforlargelanguagemodel(LLM)transformerdevelopmentperNvidia.

Additionally,NvidianotedthattheupcomingGB200withBlackwellcouldofferupto30XbetterperformancecomparedtothesamenumberofNvidiaH100sforLLMinferenceworkloadswithupto25Xlowercostandenergyconsumption.

WeconsiderTeslatobeoneoftheleadersinautonomoustechnology.Tesla’s

supervisedFSDtechnology,whichisanL2/L2+systemasitrequiresthedrivertopayattentionandbepreparedtotakeoveratalltimes,isalreadysaferthana

10July20249

humandriverintermsofaccidentspermileperdatafromTesla(Exhibit7).Tesla’sdatashowsthatin2023accidentswithFSD(whichisTesla’ssolutionthatallowsthecartodopointtopointnavigation,andworksonhighwaysandcitystreets)occurredevery4to5millionmiles,andaccidentsonAutopilot(whichincludesmorebasicfeaturesliketrafficawarecruisecontrolandlanekeeping)occurredevery5to6mnmiles(althoughthesearegenerallyhighwaymiles),comparedtoonceevery600-700KmilesfortheUSonaverage.In1Q24,accidentswithAutopilothappenedonceevery7-8millionmiles.

Exhibit7:TeslaADASsystemmilesdrivenbeforeanaccident

Milesdrivenbeforeaccident(mns)

6

5

4

3

2

1

0

201820192020202120222023

TeslaFSDBeta/SupervisedTeslaAutopilotTeslanon-AutopilotUSaverage

Source:Companydata,FHWA,NHTSA,GoldmanSachsGlobalInvestmentResearch

Whileweconsiderthisanimportantmetric(withrealworldimplications),

measuringifTesladriversgetinfeweraccidentswhileusingthetechnologywithactivedriversupervision(andtakingoverifthereisanissue)isnotthesameaswhetherthevehicleswouldbebetterand/orsaferwhenunsupervised(e.g.L3/L4capability).Wethereforebelieveitisalsohelpfultogaugethepercentageofdrives

withnointerventions,andinterventionspermile,toassesswhenTeslamayreachunsupervisedcapability(e.g.L3orL4).

WhileTesladoesn’tdiscloseinterventiondataonFSD(andinterventionsareatthe

discretionofthedriverwhichcancomplicateananalysisofthissort),somedrivers

submitinterventiondatatoTeslaFSDtrackerwhichshowsthatthepercentageofdriveswithoutaninterventionhastrendedhigherwithFSDV12(theversionwhenTeslamovedtoanend-to-endAIapproach).About70%ofdrivesonV12havebeen

no-interventionperthiscrowd-sourceddata,whichisanimprovementfromV10andV11.

GlobalAutos&IndustrialTech

10July202410

Exhibit8:TeslaFSDperformancebyversionpercrowd-sourceddata

80%

%ofdriveswithnodisengagements

70%

60%

50%

40%

30%

20%

10%

0%

12

.11.3

10.6911.4

10.11

o10.12

2/1/2022

3/1/2022

4/1/2022

5/1/2022

6/1/2022

7/1/2022

8/1/2022

9/1/2022

10/1/2022

11/1/2022

12/1/2022

1/1/2023

2/1/2023

3/1/2023

4/1/2023

5/1/2023

6/1/2023

7/1/2023

8/1/2023

9/1/2023

10/1/2023

11/1/2023

12/1/2023

1/1/2024

2/1/2024

10.10

Source:TeslaFSDTracker

Moreimportantly,whatdoesthisdatasuggestforwhenTeslacouldreachL3or

eyes-offcapability?Therearedifferentpotentialextrapolationsofthedata,includingrapidscaling,slowerbutsteadyprogress,oraneventualceilingonimprovement.Whilewerecognizethatthisdataisimperfect,itwouldimplyinouropinionthatTeslaistracking

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论