Glonce

Waymo unveils custom 5nm ASIC for self-driving compute

infrastructureThe Glonce editorsHacker News🔥62Glonce rating62Impact68Novelty62Relevance92Surprise45Players87Credibility92
Waymo unveils custom 5nm ASIC for self-driving compute

Waymo published a blog post laying out the architecture of the onboard compute system that runs its self-driving cars, and disclosed a new piece of custom silicon: a purpose-built 5nm ASIC. The chip is dedicated to the front end of the pipeline, taking raw lidar, radar and camera streams and processing, fusing and running neural networks on them in real time, including temporal denoising for better low-light perception, before handing the results to the vehicle's main ML system. Waymo says the ASIC alone delivers over 1,000 TOPS of ML performance for that front-end work. The company frames the whole system around three requirements it says are non-negotiable for driving without a human backup: responsiveness (decisions made onboard within milliseconds, with raw onboard compute power scaled 20x over the past eight years), ruggedness (hardware built to withstand vibration, shock and extreme temperatures, tied into the vehicle's liquid cooling system to hold performance from Midwest winters to Phoenix summers), and redundancy (the compute is built as two independent engines that normally run the same workload in parallel, with one taking over seamlessly if the other faults). Waymo says the design draws on more than 200 million miles of fully autonomous driving experience and describes the result as a heterogeneous system pairing the ASIC with CPUs, GPUs and other accelerators for non-ML tasks like orchestration, data movement and logging. The company named the outside partners supplying its computing components: AMD, Micron, NVIDIA, Samsung, Sandisk, Socionext and TSMC. The post does not give the ASIC a product name, a shipping timeline or a vehicle generation, and points readers to talks at the Hot Chips conference and to open roles at waymo.com/careers for more detail.

Key facts

  • Waymo introduced a purpose-built 5nm ASIC dedicated to processing raw lidar, radar and camera data in real time, delivering over 1,000 TOPS of ML performance for that front-end work alone.
  • The onboard compute system's raw processing power has scaled 20x over the past eight years.
  • The design draws on more than 200 million miles of fully autonomous driving experience.
  • Waymo named AMD, Micron, NVIDIA, Samsung, Sandisk, Socionext and TSMC as partners supplying computing components for the system.
  • The compute is built as two independent engines running parallel workloads, with automatic failover if one faults, and integrates with the vehicle's liquid cooling system to hold performance in extreme temperatures.

Why it matters

A car that drives itself with no human backup cannot tolerate a slow or unreliable brain. Waymo argues that off-the-shelf compute cannot meet the combination of millisecond-level responsiveness, physical ruggedness and built-in redundancy that fully driverless operation demands, which is why it has moved to custom silicon co-designed with its sensors and algorithms. The 5nm ASIC is the clearest evidence of that shift: a chip built specifically to turn raw sensor streams into a usable picture of the world fast enough for real-time driving decisions.

Who it affects

Waymo itself, whose robotaxi fleet depends on this compute stack. It also affects the chip and component makers named as partners, AMD, Micron, NVIDIA, Samsung, Sandisk, Socionext and TSMC, which supply pieces of the broader system around Waymo's custom ASIC. More broadly, it signals to the autonomous vehicle industry that in-house silicon, not just software, is becoming a competitive front.

How to use it

There is no consumer product or public specification to buy or download here. Waymo says more detail on its compute approach will be presented at talks at the Hot Chips conference, and it points interested engineers to open roles at waymo.com/careers.

How solid is it

The source is Waymo's own blog post, written to promote the company's engineering, not an independent teardown or third-party benchmark. The performance figures (20x scaling, over 1,000 TOPS, 200 million-plus miles) are Waymo's own claims, stated without supporting data, methodology or comparison figures.

Risks and caveats

The post does not name the ASIC, give it a shipping timeline or say which vehicle generation carries it. It gives no comparison of cost or performance against competitors or against Waymo's own prior compute generation, and no breakdown of which specific CPUs, GPUs or accelerators are paired with the ASIC beyond naming the partner companies. No individual author or engineer is credited for the post.

“While these ASICs alone deliver over 1,000 TOPS of ML performance dedicated to front-end processing and ML models, we optimize across the full stack to maximize achieved performance, especially in the low-batch regimes we often operate.”

— Waymo, official blog post