Meta News

Waymo has revealed new details about the computing architecture behind its autonomous vehicles, offering a rare look at the hardware responsible for turning enormous streams of sensor data into driving decisions within milliseconds. At the center of the system is a custom 5-nanometer ASIC capable of delivering more than 1,000 TOPS of machine-learning performance, supported by Nvidia accelerated computing and components from several major semiconductor manufacturers.

The architecture illustrates an important aspect of autonomous driving that is often overshadowed by cameras, lidar and artificial intelligence models. A robotaxi does not simply need sophisticated AI; it needs enough computing power inside the vehicle to continuously process its surroundings, predict how other road users will behave and make safety-critical decisions without depending on a remote data center.

Waymo’s latest onboard system combines its custom accelerator with Nvidia computing technology and hardware supplied by AMD, Micron, Samsung, SanDisk, Socionext and TSMC. The custom ASIC is manufactured using a 5 nm process and delivers more than 1,000 trillion operations per second for machine-learning workloads.

That computing capacity is necessary because the Waymo Driver continuously processes information from cameras, lidar and radar. Those sensors provide different representations of the environment, which must be combined to identify vehicles, pedestrians, cyclists, traffic signals and road geometry while simultaneously predicting how the scene may evolve.

The system then needs to translate those predictions into driving decisions—accelerating, braking, steering or changing lanes—with response times measured in milliseconds. Unlike a cloud AI application, latency is not merely a user-experience concern. Delayed inference inside an autonomous vehicle can become a safety issue.

Nvidia accelerated computing forms an important part of the architecture, but Waymo is not relying exclusively on general-purpose commercial AI processors. Its custom silicon provides dedicated acceleration for workloads specific to autonomous driving, while other components handle complementary computing and memory requirements.

This hybrid approach reflects a broader trend in large-scale AI systems. GPUs provide flexibility and enormous parallel computing capacity, while specialized ASICs can be optimized for workloads that are predictable enough to justify dedicated hardware.

For Waymo, that optimization matters at fleet scale. A processor that consumes less power or performs inference more efficiently can translate into lower energy requirements, less cooling and ultimately lower operating costs across thousands of vehicles.

Performance is only part of the engineering problem. Waymo’s onboard computer must operate reliably through vibration, changing temperatures and the physical stresses associated with continuous vehicle operation.

That makes the system fundamentally different from AI hardware installed inside a controlled data center. Servers operate in environments designed around stable power, temperature management and predictable networking. A robotaxi’s computer travels through summer heat, winter weather, potholes and continuous mechanical vibration while still being expected to deliver deterministic performance.

Reliability and redundancy consequently become as important as raw TOPS. An autonomous vehicle cannot simply display an error message and wait for an administrator when a component encounters a problem.

Waymo uses extensive cloud infrastructure for training, simulation and fleet development, but the immediate driving decision must happen locally. Losing connectivity cannot mean losing the ability to control the vehicle.

This separation between cloud training and edge inference is fundamental to autonomous driving. Large-scale computing infrastructure can train increasingly sophisticated models using enormous datasets and simulated environments, while optimized hardware inside the vehicle executes those models in real time.

Waymo has invested heavily in simulation as part of this process. Its World Model allows the company to expose the Waymo Driver to large numbers of virtual scenarios before encountering similar situations on public roads.

The vehicle therefore carries the result of enormous data-center workloads, but the final perception and driving loop remains an edge-computing problem.

The new computing architecture accompanies Waymo’s broader push toward its sixth-generation Waymo Driver, designed to reduce cost while supporting more vehicle platforms and a wider range of environmental conditions, including extreme winter weather.

That transition is important because autonomous driving has moved beyond demonstrating that the technology can work. The challenge is increasingly whether it can be manufactured and operated economically across large fleets.

Waymo is already moving in that direction. Its autonomous system has accumulated more than 220 million fully autonomous miles through March 2026, while the company continues expanding into additional US cities.

Its newer Ojai vehicle will be the first platform to debut the sixth-generation Waymo Driver, and the company says it is scaling manufacturing toward capacity measured in tens of thousands of vehicles annually.

Waymo’s disclosure also shows how closely the future of autonomous vehicles is tied to semiconductor engineering. Better driving models alone are insufficient if they cannot execute quickly, efficiently and reliably inside a production vehicle.

The company is effectively building a specialized AI computing platform optimized around one demanding application: interpreting the physical world and safely navigating through it in real time.

That requires cameras, radar and lidar to perceive the environment; machine-learning models to understand it; and more than 1,000 TOPS of local compute to turn that understanding into action.

The result is something increasingly similar to a mobile AI data center, except that its most important output is not text, images or recommendations. It is deciding where a multi-ton vehicle should move next—and making that decision within milliseconds.

⁠Original report at Robotics & Automation News