aimer31Aug20231pmentNewEnergyUtilities馬斯克演示FSDBetaV12“端到端”技術(shù)，推動自動駕駛加速滲透MuskDemonstratesFSDBetaVEndtoEndAITechnologyAcceleratingPenetrationofAutonomousseccom熱點速評FlashAnalysis(PleaseseeAPPENDIX1forEnglishsummary)AIC++代碼控制量約10倍 U據(jù)，車輛配備8個攝像頭，以每秒36幀的速度進(jìn)行拍攝(實際運行速度可高達(dá)每秒50幀)，系統(tǒng)則通過AI系統(tǒng)尚未支持HW4.0硬件，主要因為本次演示車型采用了“視覺輸入、控制輸出”系統(tǒng)，包括輸入視覺等傳感他汽的自動駕駛付費出行服務(wù)。上海首批獲得測試牌照的三家企業(yè)可在浦東新區(qū)內(nèi)的劃定區(qū)域開展全無人的自動企業(yè)正在引領(lǐng)自動駕駛技術(shù)應(yīng)用進(jìn)入實際落地階段。建議關(guān)注特斯拉產(chǎn)業(yè)鏈及自動駕駛產(chǎn)業(yè)鏈兩條主線的投恒I231Aug20232APPENDIX1Event:OnAugust26,Muskshowedaself-drivingtestdriveofaModelSvehicleequippedwithanFSDBetaV12.Thelivetestdrivelasted45minutes,duringwhichMuskonlytookoveronce,mainlytoconfusealeft-turnsignalwhilewaitingforaredlightstraightahead.Intherestofthetime,theFSDBetaV12systemperformedwell,achievingsmoothhandlingofcomplexroadconditionssuchasobstaclesandroundabouts.MuskannouncedthatFSDBetaV12isthefirst"end-to-end"AIautonomousdrivingsystem.Thetechnologyparadigmisbasedonrandomlymarkingpathpointsonthemap,andthesystemtakesovertheentireprocessfromstarttodestination,placingmoreemphasisonoptimizingthealgorithm.FSDBetaV12isfullytrainedbyAIneuralnetworks,anddoesn'tinvolveanytraditionalprogrammingcodetorecognizetheroad,pedestrians,etc.TheAIsystemreplacesthetraditionalcontrolmodule,andreducestheamountofC++codetobecontrolledbyafactorof10.Thesystemallowsdriverstoworklocallywithouthavingtorelyonnavigationormapdataconnectionswhentheyareunfamiliarwiththeterrain.10,000H100GPUclustersareonlinetohelpaccelerateFSDV12.ItisreportedthatTeslawillrecentlygolivewitha10,000pieceH100GPUclustertosupportthetrainingoftheFSDsystem.Thepeakarithmeticoftheclustercanreach340FP64PFLOPS,andthepeakAIarithmeticcanreach39.58INT8ExaFLOPS.Atpresent,intelligentcarsaremainlyfacedwiththechallengeoftrainingarithmetic,andtheon-linelaunchofthisclusterwillboostthefurtherdevelopmentoftheFSDBetaV12.InJuly,TeslahasalreadyputintoproductionthesupercomputerplatformDojo,andisplanningtoOctober2024toincreaseitscomputingpowerto100E,whichwillsignificantlyincreasethetrainingcomputingpower.Currently,FSDBetaV12isstillindebugging,andtheofficialreleasedateisyettobeconfirmed.ThedemonstrationvehicleadoptsHW3.0hardware,HW4.0hardwareisnotsupportedforthetimebeing.ThetrainingofFSDBetaV12reliesonalargeamountofvideotrainingdata,thevehicleisequippedwith8cameras,shootingataspeedof36framespersecond(theactualrunningspeedcanbeashighas50framespersecond),andthesystemmakesreal-timejudgmentsonthevideothroughAI.ThedrivingvehicledemonstratedinthislivebroadcastusesHW3.0hardware,whoseinferencecomputingpowerisonlyabout100watts.Currently,thesystemdoesnotyetsupportHW4.0hardware,mainlybecausethedemonstrationmodelusesa"visualinput,controloutput"system,butthetwodataarenotcompatible.MusksaidthatHW4.0willberetrainedinthefuturetoadapt.TeslaFSDwillsoonlandinthecountry,promotingtheacceleratedpenetrationoftheautopilotindustrychain.ItisreportedthatthebetaversionofFSDwillbepushedtousersinChina,andwillinitiallycollectdatain"shadowmode",andwillgraduallylaunchasmall-scaletrialaftercompletingtherelevantfeasibilityverification.InJune,MuskannouncedthathewouldopenuptheuseofFSDtechnologytootherautomakers,andinAugust,TeslaalreadysetupadatacenterinChinatorealizethelocalizationofdatastorage.Intermsofpolicy,inJuly,Beijingofficiallyopenedthecommercializationpilotof"unmannedvehicle"forintelligentnetworkedpassengercars,andenterprisesthatmeettherequirementscanprovideregularself-drivingpaidtravelservicesinthedemonstrationarea.InShanghai,thefirstthreecompaniestoobtainatestlicensecancarryoutfullyunmannedself-drivingtestsinadesignatedareainthePudongNewArea.Thecontinuousprogressofautonomousdrivingtechnology,coupledwithrelevantpolicysupport,isexpectedtofurtherpromotetheacceleratedpenetrationoftheautonomousdrivingindustrychain.Investmentadvice:WebelievethatthegraduallandingofAIinnovativetechnologiessuchasFSDBetaV12willpromotetheacceleratedpenetrationoftherelevantindustrychain,andtheautomaticdrivingenterprisesrepresentedbyTeslaareleadingtheapplicationofautomaticdrivingtechnologyintotheactuallandingstage.ItisrecommendedtopayattentiontotheTeslaindustrychainandtheinvestmentopportunitiesofthetwomainlinesoftheautonomousdrivingindustrychain.ThefocusingareasincludeARBEsolutions,largemodeltechnologyanddomaincontroller,etc.Itisrecommendedtopayattentionto[NingboTuopuGroup],[ThunderSoftwareTechnology],[BeijingJingweiHirainTechnologies],[NingboJoysonElectronic],etc.Risks:AItechnologydevelopmentlessthanexpected,marketcompetition,policylessthanexpected3下下m4k567