Waymo reveals custom 5nm ASIC and full-stack compute for fully autonomous driving

Waymo’s compute architecture is built for fully autonomous operation, where no human fallback exists. This shifts the engineering challenge from driver-assistance to deterministic, real-time decision-making. The system must process raw sensor data into driving commands within milliseconds, a requirement that off-the-shelf hardware struggles to meet consistently. By co-designing hardware, sensors, and algorithms, Waymo achieves the necessary performance while managing power, thermal, and spatial constraints of a vehicle environment.

The introduction of a custom 5nm ASIC marks a departure from reliance on general-purpose accelerators. This chip is specialized for front-end sensor processing, handling tasks like temporal denoising and sensor fusion before data reaches the core ML inference engine. With over 1,000 TOPS dedicated to these workloads, the ASIC reduces latency and improves efficiency, particularly in low-batch regimes typical of real-time driving. However, this level of customization comes with high development costs and long lead times, limiting flexibility for rapid iteration.

Redundancy and ruggedization are non-negotiable for in-vehicle compute. Waymo’s system operates as two independent engines, with seamless failover to maintain safety. The hardware is designed to withstand vibration, shock, and extreme temperatures, integrating with the vehicle’s liquid cooling to sustain performance. These requirements add complexity and cost, as every component must meet automotive-grade reliability standards. The trade-off is a system that can operate in diverse environments, from freezing winters to desert heat, without degradation.

The compute stack balances heterogeneous components, pairing the custom ASIC with CPUs, GPUs, and other accelerators. This hybrid approach maximizes efficiency for both ML and non-ML tasks, such as orchestration and logging. By optimizing dynamic workloads, Waymo reduces “pixels-to-actuation” latency, improving responsiveness in high-density environments. However, the system’s performance is tightly coupled to its software stack, meaning optimizations are specific to Waymo’s algorithms and sensor suite. Porting this architecture to other autonomous platforms would require significant re-engineering.