20230706_光大证券_汽车和汽车零部件行业AI大模型应用于汽车智能驾驶梳理_32页.pdf
2023 7 6 CFA S0930515090002 AI L2 L3 2D+CNN BEV+Transformer 1 2 3+L3 L3 1 2022+BEV+Transformer 1+2+L3 L2/L2+L3=+1 FSD+Dojo 2 BEV+Transformer Occupancy+Transformer 3+1 2 YWFUzQtOmPnMtNnOrMpQpOaQdN8OpNmMsQmPiNnNsNkPrRrO8OnNuMwMrMvNvPpMmR 2 3 1 2 end-to-end+3D 4 L2 L3 L3 L3 360+L2 L3 L0 L1 L2 L3 L4 L5+5+L3 2022 2D+CNN ChatGPT AI BEV+Transformer 2021+2023 BEV+Transformer vs.UniAD+Transformer 1+L3 L3 2+Yihan Hu Planning-oriented Autonomous Driving AI DAY 6 BEV+Transformer BEV+Transformer 1 BEV 2D BEV 3D 2 BEV BEV 3D+3 BEV UniBEV BEV 3D+V2X BEV UniBEV 7 BEV+Transformer BEV BEV 2D+CNN 2D CNN 3D BEV Transformer 2D BEV Queries 3D BEV 1 BEV 2 BEV 3 BEV BEV vs.2D+CNN+BEV+BEV Single-Camera FrontendLidar FrontendOther Sensor FrontendsCross-Stream AlignmentCross-Modality AlignmentLearned spatial&Temporal AggregationLow-level-PhysicsSematic-level Entity ExtractionFeature MapsImplicit InformationStructure-level Concepts,Relations,BehaviorsCamerasLiDARIMUGPSWheel odometryEdge Data MiningCloud-based World ReconstructionQuality InspectionCloud-based Offline Perception 8 Transformer vs.CNN Transformer+Corner-case+Zhiqi Li BEVFormer:Learning Bird s Eye-View Representation from Multi-Camera Images via Spatiotemporal Transformers BEV+Transformer BEV+Transformer Transformer 9+1 2 3 vs./+BEV+Transformer AI DAY Occupancy 10 11+1 2016 Mobileye 2 2016-2017 3 2018 FSD 1 2018 2 2021 3 2021-2022 4+Dojo 2016 2016-2017 2018-2020 2021 2022 Autopilot1.0 Autopilot2.0 Autopilot 3.0 FSD/NOA 2D+CNN HydraNets 2D+CNN BEV+Transformer Occupancy+Transformer+HW1.0 HW2.0+HW2.5 2017 2.0 HW2.5+HW3.0 HW3.0 HW3.0 2023/3 Model S/X HW 4.0 Mobileye EyeQ3 DRIVE PX 2 FSD*2/TOPS 0.256 12 12 144 144 NA NA NA A100 Dojo NA NA NA 1.8ExaFlops 12 HW1.0 HW3.0 HW3.0 FSD+FSD HW4.0 2023/3 Model S/X 2 FSD2.0 300-500TOPS vs.Orin 254TOPS L3-L4 L3 100-200TOPS L4 400-600TOPS HW1.0-HW3.0 HW1.0 HW2.0 HW2.5 HW3.0 1 Camera 35 1/Camera 50 1/Camera 120 1 0 Camera 90 2 0 Camera 60 2 Radar 1 160m Radar 1 170m UUS 12 5m UUS 12 8m Mobileye EyeQ3 1 Nvidia Parker SoC 1 Nvidia Pascal GPU 1 TriCore MCU 1 Nvidia Parker SoC 2 Nvidia Pascal GPU 1 TriCore MCU 1 FSD 2 0.256TOPS 12TOPS 12TOPS 144TOPS 72TOPS ROM 256 6GB 8GB 8GB 2 Flash NA NA NA 4GB 2 1 40 40 420 36 110 110 2300 25W 250W 40W 300W 220W 2014 Model S/X 2016 Model S/X 2017 Model S/X/3 2019 Model S/X/3 3 HW2.5 13 1 2+vs.L3-L4 1 2021 Dojo A100 2H23E 2 3/AI DAY 2D+CNN BEV+Transformer Dojo 2021 2021 NA NA 720 8 NVIDIA A100 Tensor Core GPU 5,760 GPU D1 NA 2TB/s D1 10TB/s on-chip bandwidth vs.Stratix 10MX 1TB/s 4TB/s off-chip bandwidth NA 100 36 10 1.5PB 30PB NA NVIDIA A100 Tensor Core 19.5 TFLOPS FP32 1.8 exaflops Dojo 3,000 D1 D1 22.6 TFLOPS FP32 Dojo 1.1 Exaflops 7 Dojo 8Exaflops 14 Occupancy Network 36 2016Software 1.0 2D+CNN cls reg640 480 1 640 480 4ResNet/RegNetRaw2018Autopilot 4.0 HydraNets 2D+CNN clsDecoder TrunkTask 1RegNet+BiFPNRawregTask 2attr cls regDecoder TrunkRadarMulti-Scale Features2021Software 2.0 BEV+Transformer clsDecoder TrunkTask 1RegNet+BiFPNRawregTask 2attr cls regDecoder TrunkMulti-ScaleFeaturesRaw RawRectify Rectify RectifyRegNet+BiFPNRegNet+BiFPNMulti-ScaleFeaturesMulti-ScaleFeaturesMulti-Camera Fusion&BEVTransformer2022FSD Beta Occupancy+Transformer clsDecoder TrunkTask 1RegNet+BiFPNRawregTask 2attr cls regDecoder TrunkMulti-ScaleFeaturesRaw RawRectify Rectify RectifyRegNet+BiFPNRegNet+BiFPNMulti-ScaleFeaturesMulti-ScaleFeaturesMulti-Camera Fusion&BEVTransformerVideo Features Queue/Video Module 2016 1 2018 2D+CNN 2 2018-2020 HydraNet RegNet+BiFPN 3 2021+BEV+Transformer Dojo 4 2022 BEV Occupancy 1 2+15+AI DAY+1+corner case 2+16 17 L3 2022 BEV+Transformer 2022/10 2023/4 2023/4/2H23E+L3 L2+2020 2021 2022 2023 FSDFSD BetaFSD Betav8FSD Beta v9.0FSD Betav10.0FSD Betav11 FSD Betav11FSD Betav11.3.2NOP NOP+NAD NOP+AD Max 1 10 12 1 9 12 Hi NCANCA NCAADS2.0 AD PRO NOAAD MAX 3.0 XNGP4.2.0 NGP XNet NGP 4 7 6 11 12 9 10 8 1 4 6 2 3 7 18 ADS2.0+1 ADS 2.0 NOP+/2 1+/GOD Occupancy+Transformer/BEV+Transformer 2 2023 30-35+AD Max+42 NOP+NOA XNGP ADS 2.0 L2+L2+L2+L2+/Y 70km/h 80km/h&60km/h 60km/h 19 ADS2.0 42 NOP+NOA XNGP ADS 2.0 80km/h 80km/h 80km/h XNGP 70km/h 70km/h 80km/h&60km/h 60km/h 20 ADS2.0 IT 2023/6/30 2022 ADS2.0 XNGP NOA NOP+*8*1*12*11*3*12*1*11*5*12*2*11*1*12*1*11*5*12*1 HW4.0,144TOPS 610 200TOPS Orin 508TOPS Orin 508TOPS Orin 1,016TOPS A100 1.8 exaflops Dojo 1.1exaflops 7 Dojo 8exaflops ADCSC 400-800TOPS 600pflops NVIDIA HGX A100 8-GPU NVIDIA Mellanox InfiniBand ConnectX-6 Petal Map NA+Occupancy+Transformer+GOD Occupancy+Transformer Xnet(BEV+Transformer+)Occupancy+Transformer+BEV+Transformer 64,000 EAP 32,000 ADS2.0 36,000 2023/6/1-12/31 18,000 7,200 720 6 ADS2.0 G9 P7i Max 2 AD MAX 4 380 21 ADS2.0 IT 2023/6/30 2022 ADS2.0 XNGP NOA NOP+2023/6 FSD Beta 3.06 NA 2022/6 NGP 62%2023/3/23 5.5 NOA 1 2023/4/25 100 8.2 NOP 3.1 NOP+2730 NOA FSD beta 2Q23/3Q23 15 4Q23 45 2023/6 2H23 2023/1 Beta 2023/7/1 NOA NGP NOA APA AVP VPA S-APA L2+L2+L2+L2+L2+79/107/84/19/9 298 70 1,000 90/153/154/76/39/6 518 152 65/169/128/112/24/11 509 154 64/115/135/39/15/1 369 98 22 23 BEV+Transformer+1 2+24 corner cases+1 trigger 2+3+4+5+25 L3 L2/L2+L3=1 L3+vs.L3-L4 2+/+/+2H23E TSLA.O LI.O XPEV.N NIO.N 26 vs.L2/L2+L3/+NOA+AVP SoC 1 2 3+002920.SZ 688326.SH 603786.SH OTA/ECU 27/1 L2/L2+ACC ASIL A QM vs.L3 ASIL C ASD 2+/3 1 2+2025E 40%-50%10%one box 603596.SH 1316.HK Review on Automobile Steering-by-wire System Development/One Box Two Box iBooster+ESP EPS SBW 28 29 1 2 1 2 3 4 5 6 NOA 7 8 1 2 3 CFA S0930515090002 021-52523876 6-12 15%6-12 5%15%6-12-5%5%6-12 5%15%6-12 15%A 300 500 1996 500 鹏华基金
