EV ADAS/AV Compute Platforms

ADAS (Advanced Driver-Assistance Systems) and AV (Autonomous Vehicle) compute is the hardware that runs perception, sensor fusion, planning, and control decision workloads in real time. In most vehicles, this compute is primarily an inference platform: it runs trained models and deterministic algorithms on-vehicle to make driving decisions without relying on cloud latency.

Key terms

• Inference: running trained models on the vehicle to interpret sensor data and generate decisions
• SoC (System-on-Chip): integrated compute chip combining CPU cores and accelerators (GPU/NPU)
• NPU (Neural Processing Unit): AI accelerator optimized for matrix operations
• Safety MCU: deterministic safety controller supervising compute and outputs

Inference platform vs central compute

Many OEMs use “inference computer” and “central ADAS compute” interchangeably. The practical distinction is architectural:

• Inference platform: the compute module(s) responsible for real-time ADAS/AV decision workloads
• Central compute: a broader architecture trend where fewer computers run multiple domains; ADAS inference may be one of those domains

Compute term	What it usually means	Typical scope	Example contents
Inference computer / inference platform	Vehicle-side compute that runs real-time perception and planning	ADAS/AV domain	SoC(s), safety supervision, DRAM, storage, Ethernet
ADAS computer	OEM label for the ADAS compute module	ADAS/AV domain	Same as inference platform in most contexts
Domain controller (ADAS DCU)	Domain compute that may include inference workloads	ADAS plus gateway/supervision	SoC or high-end MCU, Ethernet, CAN gateway, safety
Central compute	Fewer computers running multiple vehicle domains	Multi-domain vehicle platform	SoC(s) + virtualization + networking fabric + safety supervision

Compute architecture

ADAS/AV inference platforms are built from a small set of repeatable hardware blocks.

Block	What it does	Why it matters
SoC (CPU + GPU/NPU)	Runs perception, fusion, planning, and runtime systems	Primary performance driver for ADAS capabilities
AI accelerator (NPU / dedicated accelerator)	High-throughput model inference	Determines real-time model capacity under power constraints
Safety supervision (safety MCU / safety islands)	Monitors health and validates outputs; triggers safe state	Required for functional safety and higher autonomy
DRAM	Runtime memory for perception pipelines and models	Capacity and bandwidth can bottleneck performance
Storage (flash / NVMe)	Logs, maps, firmware, local buffering	Enables data capture and robust OTA update staging
Networking (Ethernet)	Moves sensor data to compute and commands back to controllers	Scales with camera/LiDAR payloads and topology
Power delivery	Provides stable rails and transient handling	Compute stability affects safety and reliability
Thermal solution	Removes heat from SoCs and memory	High-end ADAS compute can become a primary thermal load

Compute scale by autonomy level

Compute requirements scale with sensor count, model complexity, and redundancy requirements.

Autonomy level (informal)	Typical compute pattern	Redundancy expectation	Network implication
L2 (ADAS)	Single ADAS SoC module or integrated ADAS ECU	Fail-safe; limited redundancy	Ethernet begins for cameras; CAN remains dominant for control
L2+ / L3 (conditional)	Higher-end SoC with more sensors and compute headroom	More supervision; increasing redundancy	Ethernet backbone expands; gateways bridge legacy buses
L4 (high autonomy)	Multi-SoC or redundant compute modules; higher storage and data logging	Fail-operational; physical redundancy expected	High-speed Ethernet fabrics; 10G-T1 becomes relevant

Safety supervision and command integrity

Safety supervision is the hardware that prevents unsafe commands from propagating when compute faults occur.

• Safety islands inside SoCs: dedicated safety logic within the main compute chip
• External safety MCU: independent controller that monitors health, cross-checks outputs, and triggers safe-state transitions
• Independent power rails and watchdog circuits: keep supervision operating through transients

Memory, storage, and data logging

ADAS compute is increasingly data-centric. Even when autonomy is not fully deployed, vehicles log edge cases for analysis and improvement.

• DRAM: impacts real-time perception throughput
• Persistent storage: impacts logging depth and OTA staging reliability
• Write endurance and thermal stability become practical constraints

Networking interfaces to the IVN

ADAS compute platforms are typically Ethernet-rich nodes.

• Multiple Ethernet ports for sensor ingestion and zonal uplinks
• Gateway interfaces to CAN/CAN-FD and LIN via DCUs or ZCUs (architecture dependent)
• Time synchronization support becomes more important as sensor fusion scales

Power and thermal constraints

Compute capability is limited by the vehicle power and thermal envelope.

• Power: low-voltage rails (12 V or 48 V) must support sustained compute draw plus transient loads
• Thermal: air cooling may be sufficient at lower compute levels; liquid cooling is used in higher-end designs
• Reliability: automotive temperature ranges and vibration drive packaging and connector requirements

Supply-chain notes

ADAS/AV compute is a high-value concentration point in the vehicle BOM.

• Automotive-grade SoCs and accelerators: constrained by advanced semiconductor nodes and packaging
• DRAM and flash: capacity and bandwidth trends track sensor content
• Ethernet switches and PHYs: required for high-rate sensor ingestion and zonal backbones
• Thermal hardware: cold plates, heat sinks, and thermal interface materials scale with compute
• Functional safety components: safety MCUs and safety islands are key differentiators

Inference compute platforms

OEMs choose from a few common platform categories. Specific model names and branding change frequently, but the underlying supply-chain pattern is stable.

• Integrated ADAS SoC platforms: CPU plus GPU/NPU on a single chip, packaged as an automotive compute module
• Multi-chip accelerator platforms: SoC plus one or more accelerators, sometimes with dedicated safety controllers
• In-house platforms: OEM-designed inference chips and boards integrated tightly with vehicle OS and sensor suite

Tesla FSD, Mobileye EyeQ, NVIDIA DRIVE, and Qualcomm Snapdragon Ride are the most popular inference compute chips.

ADAS for cars

Car brand	ADAS system	Inference compute platform
Aion	ADiGO	NVIDIA DRIVE Orin
Aito	Qiankun ADAS (Huawei)	Huawei ADS SoC
Arcfox	α-Pilot ADS	Huawei Kirin 990A
Audi	Audi Pre Sense	Infineon
AVATR	Qiankun ADAS (Huawei)	Huawei ADS SoC
Baojun	Chengxing	Zhuoyu 7V+32
BMW	BMW Personal Pilot L3	Qualcomm Snapdragon Ride
BYD	DiPilot	NVIDIA DRIVE Orin/Thor
Chevrolet	UltraCruise	Snapdragon, Infineon
Denza	DiPilot	NVIDIA Drive Orin
Ford	BlueCruise	Intel Mobileye EyeQ
Geely	G-Pilot	NVIDIA DRIVE Orin | Ecarx
GMC	Super Cruise	Qualcomm Snapdragon Ride
GWM		Qualcomm Snapdragon Ride
HiPhi	HiPhi Pilot	NVIDIA DRIVE Orin
Huawei	Huawei ADAS	Ascend AI
Hyper		NVIDIA DRIVE Thor
Hyundai	SmartSense	NVIDIA DRIVE Orin
IM	UNP	NVIDIA Drive Orin
Jaguar	InControl	NVIDIA Drive Orin
Land Rover	InControl	NVIDIA Drive Orin
Leapmotor	Leapmotor Pilot	NVIDIA Drive Orin
Li Auto	NOA	NVIDIA DRIVE Thor
Lotus		NVIDIA DRIVE Orin
Lucid	DreamDrive	NVIDIA DRIVE Orin
Luxeed	Qiankun ADAS (Huawei)	Huawei ADS SoC
Lynk & Co	G-Pilot	NVIDIA Drive Orin
Mazda	iActivsense	Qualcomm SA8155P
Mercedes-Benz	Drive Pilot	NVIDIA DRIVE Orin
Neta	Hozon Pilot	NVIDIA Drive Orin
NIO	NOP+	NVIDIA DRIVE Orin
Nissan	ProPILOT	Renesas R-Car
Polestar	ADAS	NVIDIA DRIVE Orin
Porsche	InnoDrive	Intel Mobileye EyeQ
Renault	Active Driver Assist	Qualcomm Snapdragon Ride
Rivian	Rivian Driver+	NVIDIA DRIVE Orin
SAIC	Navigation on Pilot	Horizon Robotics Journey 6
Smart	SuperVision	Intel Mobileye EyeQ
Stelato	Qiankun ADAS (Huawei)	Huawei ADS SoC
Stellantis	STLA AutoDrive	Qualcomm Snapdragon Ride
Subaru	EyeSight	Xilinx Zynq UltraScale+
Tata		Qualcomm Snapdragon Ride
Tesla	Autopilot	Tesla FSD
Toyota	Safety Sense	NVIDIA DRIVE PX
Volvo		NVIDIA DRIVE Orin
Voyah	Voyah Vpilot	Qualcomm Snapdragon Ride
VW	Travel Cruise	Qualcomm Snapdragon Ride
Wuling		Horizon Robotics Journey
Xiaomi	Xiaomi Pilot (XP)	NVIDIA DRIVE Orin
Xpeng	XNGP	NVIDIA DRIVE Orin/Thor
Yangwang		NVIDIA DRIVE Orin
Zeekr	Huohan 2.0+	NVIDIA DRIVE Thor

ADAS for long-haul trucks

Heavy-duty Class 7-8 EVs and hybrids. Require redundancy, sensor fusion, and specialized highway autonomy. Vendors include NVIDIA, Aurora, Plus.ai.

Trucks Brand	ADAS system	Chip platform
MAN	Plus SuperDrive	NVIDIA DRIVE Thor
Scania	Plus SuperDrive	NVIDIA DRIVE Thor
Tesla Semi	Autopilot	Tesla FSD

Robotaxi AV platforms

Urban autonomous fleets with dense sensor arrays and high compute density. Examples: Waymo, Cruise, Baidu Apollo, Pony.ai. Autonomous platforms for cars are still largely in testing phase, with various technology companies offering their "bolt-on" technology to existing car models. The exceptions are Tesla, DiDi, and Zoox, where the technology is integrated during production.

Robotaxi brand	Autonomous driver	Chip platform
Avride	Avride Driver
Baidu	Apollo Go	Kunlun AI
DiDi		NVIDIA DRIVE Orin
Motional	Motional	NVIDIA DRIVE Orin
Tesla	Cybercab	Tesla FSD
Waymo	Waymo Driver	NVIDIA
WeRide	WeRide One	NVIDIA DRIVE Thor
Zoox	Zoox ADP	NVIDIA DRIVE Orin

Robotruck AV platforms

Highway freight haulers with autonomy stacks tuned for long corridors and platooning. Autonomous platforms for trucks is still largely in testing phase, with various technology companies offering their "bolt-on" technology to existing truck brands. The exceptions are Tesla Semi and Gatik Carrier, where the technology is integrated.

Tech brand	Autonomous driver	Chip platform
Aurora	Aurora Driver	NVIDIA DRIVE Orin
Einride	Einride autonomy	NVIDIA DRIVE Orin
Gatik Carrier	Gatik autonomy	NVIDIA DRIVE Orin
Kodiak	Kodiak Driver	NVIDIA DRIVE Orin
Plus	Plus.ai SuperDrive	NVIDIA DRIVE Thor
Pony	PonyPilot	NVIDIA DRIVE Orin
Torc Robotics	Torc autonomy	NVIDIA DRIVE Orin
Waabi	Waabi Driver	NVIDIA DRIVE Thor
Wayve	Wayve AI Driver	NVIDIA DRIVE Orin

Robovan AV platforms

Last-mile delivery vans with mid-scale inference stacks. Balance cost efficiency with safety. Examples: Nuro, Udelv.

Robovan	Autonomous driver	Chip platform
Neolix	Neolix autonomy
Nuro	Nuro Driver	NVIDIA Drive Orin
Tesla	Autopilot	Tesla FSD
Udelv	Udely Transporter
WeRide	WeRide One	NVIDIA DRIVE Thor

Autonomous tech platforms

Generalized autonomy stacks licensed across multiple form factors.

Brand	Product	Chip platform
Aurora	Aurora Driver	NVIDIA DRIVE Orin
Avride	Avride Driver
Baidu	Apollo Go	Kunlun AI
Cartken
Coco
DiDi		NVIDIA DRIVE Orin
Einride	Einride autonomy	NVIDIA DRIVE Orin
Gatik Carrier	Gatik autonomy	NVIDIA DRIVE Orin
Kodiak	Kodiak Driver	NVIDIA DRIVE Orin
Motional	Motional	NVIDIA DRIVE Orin
Neolix	Neolix autonomy
Nuro	Nuro Driver	NVIDIA Drive Orin
Plus	Plus.ai SuperDrive	NVIDIA DRIVE Thor
Pony	PonyPilot	NVIDIA DRIVE Orin
Serve
Tesla	Cybercab	Tesla FSD
Tesla	Autopilot	Tesla FSD
Torc Robotics	Torc autonomy	NVIDIA DRIVE Orin
Udelv	Udely Transporter
Waabi	Waabi Driver	NVIDIA DRIVE Thor
Waymo	Waymo Driver	NVIDIA
Wayve	Wayve AI Driver	NVIDIA DRIVE Orin
WeRide	WeRide One	NVIDIA DRIVE Thor
WeRide	WeRide One	NVIDIA DRIVE Thor
Zoox	Zoox ADP	NVIDIA DRIVE Orin