Waymo Custom Chip: 5nm ASIC for Lidar and Cameras

What Waymo's 5nm ASIC does with raw sensor data

The chip takes in raw data from the car's lidar, radar and cameras and prepares it before it reaches what the post calls the core ML brain. An ASIC (application-specific integrated circuit) is a chip designed for one task instead of general use. Waymo introduced this one in a post dated August 20, 2026, written by Satish Jeyachandran, Vice President of Engineering, and Daniel Rosenband, Compute Lead.

The ASIC does not drive the car. Waymo places it ahead of that core ML brain. It handles the large flow of raw data arriving from the sensors, fuses it, and runs neural networks on it in real time.

The announced ASIC at a glance

Property	Detail
Type	Purpose-built ASIC
Process node	5nm (foundry undisclosed)
Role	Prepare raw sensor data and run ML models on it ahead of the core ML brain
Sensors	Lidar, radar and cameras
Cameras	13 high-resolution cameras at once in the latest system
Status	One of several custom components in development; the post does not describe the others

Does 1,000 TOPS cover the whole compute system?

No. TOPS stands for trillions of operations per second, a count of how much machine-learning arithmetic a chip can perform. Waymo says the ASICs alone deliver over 1,000 TOPS, dedicated to front-end processing and its ML models. The post gives no figure for the compute system as a whole.

Waymo adds that it optimizes across the full stack to raise the performance actually achieved, particularly in the low-batch regimes it often runs in. That is a separate claim from the 1,000 TOPS, which the post ties to the front-end chips alone.

How the front end handles lidar, radar and cameras

Waymo describes the front end as extraction followed by sensor fusion. Specialized accelerators pull out the information that matters, including noise removal across time, and a purpose-built inference engine then runs sensor fusion models.

Work the front end takes on

Step	What the post says
Extraction	Specialized accelerators pull critical information out of the raw lidar, radar and camera streams
Temporal denoising	Noise removal over time, part of the extraction work, aimed at low-light perception
Sensor fusion	A purpose-built inference engine runs the fusion models, which merge the different sensor types into one view of the scene

Waymo says this design gives greater efficiency without sacrificing fidelity. It also says that developing silicon, sensors and algorithms side by side lets it push sensor fidelity, bandwidth efficiency and quantization (storing numbers in fewer bits). The aim is to run models from sparse convolutions to dense transformers.

How does temporal denoising help in low light?

The post says the accelerators include temporal denoising for better low-light perception, and gives no mechanism beyond that. What it does state is what the latest system delivers overall: it processes 13 high-resolution cameras simultaneously and in real time, and shows environmental details that traditional cameras miss. Waymo illustrates this with a left and right image comparison.

Redundancy, latency and ruggedization in compute hardware

Waymo says it built its compute around three requirements it calls non-negotiable: responsive, ruggedized and redundant. They draw on more than 200 million miles of fully autonomous experience. The bar is high because the system does the whole job of driving with no human backup, unlike driver-assist computing.

The three stated requirements

Requirement	What the post says
Responsive	Everything runs onboard and decides within milliseconds. Raw compute power scaled 20x in eight years
Ruggedized	Built for constant vibration, shock and extreme temperatures, and tied into the vehicle's liquid cooling for freezing Midwest winters and the heat of Phoenix
Redundant	Two independent engines that normally run as one unit. If one has a fault, the other takes over

Because no human can take over, the post says, safety and redundancy are built in from the start. On latency, Waymo says it lowered the delay from what the sensors see to what the car does by tuning dynamic workloads at every percentile. It does not give the new delay in milliseconds.

The post uses the words "ASICs" and "two independent engines" in different paragraphs and does not link them. Reading the two engines as two ASICs goes beyond the text.

Waymo says the move from off-the-shelf parts to a custom system required rethinking its physical and architectural design, and the benefits it names concern the vehicle. A highly integrated system delivers a lot of processing power without hurting the rider experience. It saves battery, leaves ample trunk space and runs silently.

The architecture is ML-primary, meaning it is built first to run neural networks at minimal latency. Tasks that are not ML, such as orchestration, data movement and logging, are handled by pairing Waymo's ML hardware with CPUs, GPUs and accelerators, which leaves more time for ML computation. Waymo calls the result a balanced, heterogeneous system, one made of different kinds of processors.

Custom ASIC alongside Nvidia and six other named partners

The post does not announce a move away from suppliers. Next to its custom silicon, Waymo says it works closely with industry leaders whose computing solutions give it a foundation to scale efficiently. It names seven: AMD, Micron, Nvidia, Samsung, Sandisk, Socionext and TSMC.

Partners named in the post

Company	Component the post assigns to it
AMD	None named
Micron	None named
Nvidia	None named
Samsung	None named
Sandisk	None named
Socionext	None named
TSMC	None named

The post calls the ASIC one of several custom components and does not say any partner is being replaced. It also never says the chip reduces Waymo's reliance on a supplier.

Does the post say which partner supplies which part?

No. The post introduces the seven only as "a number of partners like" these companies, without matching any of them to a component. Reports that tie TSMC to the 5nm ASIC therefore go beyond what the post says. TSMC is one name on the list, and the post does not state who manufactures the chip.

In short, the Waymo custom chip handles the sensor front end. It does not drive the car, and the post does not describe it displacing any supplier. The post closes by calling itself a glimpse of what is to come, and it points readers to Waymo's talks at Hot Chips without giving a date or session.