The chips and sensors behind China’s assisted driving

Behind every NOA system there is a computer that has to process the video of seven to eleven cameras, often a lidar and several radars, and output a trajectory many times per second, inside a box that fits behind the dashboard and survives from minus 40 to plus 85 °C. This article looks at that hardware: the chips Chinese carmakers use, why several of them now design their own, and how sensors have become cheap enough for small cars.

What the computer has to do

A driver assistance controller (the “domain controller”) is a system on chip with three main parts: CPU cores for the operating system, the safety logic and the planning code; image signal processors that turn the raw output of the camera sensors into usable frames; and a neural processing unit, the part measured in TOPS (trillions of operations per second, usually on 8-bit integers). Next to it there is often a separate safety microcontroller, certified to ASIL D, which checks that the main chip is alive and can bring the car to a controlled stop if it is not.

TOPS figures are the headline number in every launch, but they are only a rough guide. What a chip can actually run depends on memory bandwidth (large models are often limited by how fast weights can be moved, not by arithmetic), on which operations the NPU supports natively, and on the software tools. A transformer based end to end model that runs well on one chip can be slow on another with similar TOPS.

Three tiers of hardware

BYD’s God’s Eye range shows how the market is segmented. The company offers three levels of the system, built on platforms of roughly 100, 300 and 600 TOPS (CarNewsChina):

• Entry level, about 100 TOPS, cameras and radars only: highway NOA and parking, in cars that cost less than 10,000 dollars.
• Middle, a few hundred TOPS, often one lidar: city NOA in mid priced cars.
• Top, 500 TOPS and more, sometimes with two chips: full city NOA with the largest models, in premium cars.

Nvidia, and the move away from it

For several years the reference chip for advanced systems was Nvidia’s Drive Orin, rated at 254 TOPS, often fitted in pairs. Its successor, Drive Thor, has been entering production cars since 2025. Nvidia’s strength is not only the silicon but the software stack, which lets carmakers move models from the training cluster to the car with little rework.

Chinese companies have been building alternatives, for reasons of cost, control and supply security:

• Horizon Robotics sells its Journey chips to many carmakers. The Journey 6 family spans from entry level parts to the J6P, at about 560 TOPS, which entered mass production in late 2025 together with Horizon’s own HSD software (Wikipedia). The company has announced a J7 for 2027 (ChinaEVHome).
• NIO designed the Shenji NX9031, a 5 nm chip that it says replaces four Orin X. It went into the ET9 flagship in 2025 and then into the rest of the range (Wikipedia).
• XPeng developed the Turing chip, designed around the needs of its own large models, and uses it in its newer cars instead of Nvidia parts (KrASIA).
• Huawei uses its own Ascend based computing platforms for ADS.

Designing a chip only makes sense with high volumes and a model architecture you control. A carmaker that designs the network and the silicon together can drop the operations it does not need and add memory where its models are starved. That is the same logic Apple followed with its phones.

Sensors: cameras first, lidar when the budget allows

A typical city NOA car has 7 to 11 cameras (front cameras with different focal lengths, side cameras, rear camera and surround view cameras used for parking), 3 to 5 radars and 12 ultrasonic sensors. The front cameras have moved to 8 megapixels, which lets the system read traffic lights and signs at a distance and spot small objects earlier.

Lidar is where China changed the economics. Chinese suppliers such as Hesai and RoboSense have pushed prices of automotive lidar down to a few hundred dollars for high volume models, a fraction of what the first units cost. The result is that a roof or windscreen lidar now appears on cars in the 150,000 to 200,000 yuan range. The debate between “cameras only” (Tesla, and the entry tiers of most Chinese brands) and “cameras plus lidar” is no longer only about cost. Lidar gives a direct distance measurement that does not depend on lighting, which is valuable at night and with unusual obstacles; cameras alone keep the system cheaper and simpler. Most Chinese carmakers now offer both, depending on the price of the car.

Power and heat

A detail that is easy to forget: a domain controller running a large model can draw 100 to 200 W, sometimes more, continuously. In an electric car that is a small but measurable cost in range during city driving, and the heat must be removed, often with liquid cooling shared with other electronics. Better efficiency (performance per watt, not only peak TOPS) is one of the reasons for the move to 5 nm and smaller processes.

Where it goes next

The models are growing quickly. XPeng and Li Auto have spoken of on-board driving models with billions of parameters, and both use their data centres to train much larger models that are then distilled into the version that runs in the car. The hardware cycle in cars is slow, three to five years from design to production, so the chips being designed now must be ready for models that do not exist yet. That uncertainty is exactly why some carmakers want to control their own silicon, and why others prefer to stay with a supplier that takes the risk for them.

For the software side of the story, see what NOA is and how reinforcement learning is used to train it.