Waymo reveals the custom silicon powering its robotaxi fleet

Waymo has opened up about the computational hardware installed in its robotaxis, unveiling specifics about its custom silicon and the network of suppliers behind its system. The company operates approximately 4,000 vehicles across more than 10 cities, performing roughly 500,000 paid trips weekly.

The Architecture Behind Autonomous Driving#

Waymo's vehicles are equipped with multiple cameras, lidar sensors, and radar units that generate enormous volumes of raw sensor data. This information flows through a suite of processors housed in the vehicle's trunk that must interpret the surroundings instantaneously and convert that understanding into driving decisions—all within milliseconds.

The company's compute stack adheres to three foundational design principles: it must react swiftly to sensor inputs with minimal delay, withstand the physical demands of road travel and temperature fluctuations, and incorporate redundancy so that a backup system activates if the primary one fails. Over the past eight years, Waymo has increased its compute capacity 20-fold. The system must also remain compact enough not to monopolize trunk space needed for passengers and luggage, and operate quietly.

Custom Silicon for Sensor Processing#

Central to Waymo's approach is a proprietary 5-nanometer ASIC—a chip engineered specifically for autonomous driving. Rather than channeling raw, unprocessed sensor streams directly to the vehicle's main processing unit, this front-end processor cleans and fuses incoming data while executing preliminary AI evaluations in real time. The chip delivers 1,000 TOPS (trillion operations per second) of processing throughput.

This architecture allows Waymo to handle information from 13 high-resolution cameras simultaneously while maintaining low-latency responses. The company claims its system achieves superior low-light perception compared to traditional camera systems, enabling the vehicle to detect obstacles and road features in darkness that conventional approaches often overlook.

A Heterogeneous Computing Stack#

Waymo's executives describe their approach as "ML-primary," emphasizing machine learning at the core while pairing it with high-performance CPUs, GPUs, and specialized accelerators. This balanced configuration manages non-machine-learning tasks—orchestration, data movement, and logging—while reserving maximum processing cycles for AI computation.

The company does not manufacture all components internally. Instead, it partners with established suppliers including to provide specialized components. Waymo concentrates its internal engineering on real-time sensor fusion and edge machine learning, leveraging global supply-chain expertise for commoditized hardware.