Waymo Unveils Custom 5nm ASIC to Power Real-Time Autonomous Driving with 1,000 TOPS Performance

Waymo Unveils Custom 5nm ASIC to Power Real-Time Autonomous Driving with 1,000 TOPS Performance The ride-sharing firm's new processor handles sensor fusion & neural network inference with ultra-low latency onboard. At a Glance - Waymo's new 5nm ASIC delivers 1,000 TOPS to process lidar, radar and camera data in real time. - The custom chip reduces temporal noise in low-light conditions to enhance object perception. - Dual redundant compute systems ensure seamless failover if one unit experiences a fault during operation. Autonomous ride-sharing company Waymo has found that delivering the compute power necessary to deliver fully autonomous vehicles that deliver passengers to destinations with no human intervention requires compute power demanded a new purpose-built 5nm application-specific integrated circuit (ASIC). Writing for the company’s blog, Satish Jeyachandran, Vice President of Engineering, and Daniel Rosenband, Compute Lead, introduced Waymo’s latest processing hardware, which is designed to process the avalanche of incoming raw data before it reaches the ride-share car’s core machine learning compute system. They describe this new chip as a “specialized machine learning powerhouse engineered exclusively to process, fuse, and run advanced neural networks on raw sensor data in real time.” That sounds impressive, as does the 1,000 TOPS ASIC system's ability to extract critical information from the raw lidar, radar, and camera data streams. The company specifically calls out the chip’s ability to reduce the temporal noise in low-light images to improve the perception of objects. Waymo’s purpose-built 5nm ASIC. WAYMO The authors describe a trio of hardware requirements they say are non-negotiable: Responsive: To make real-time driving decisions, the autonomous system operates entirely onboard, constantly processing decisions within milliseconds. We have engineered our stack for ultra-low latency, minimizing the delay from first pixel to action. Ruggedized: The hardware operates under constant vibration, shock, and extreme temperatures. Redundant: While they normally operate as one unit running full parallel workloads, if one compute system experiences a fault, the other seamlessly takes over. Waymo had advanced its own in-house silicon’s compute power by a factor of 20x in eight years, and taps into its cars’ liquid cooling system to keep those chips in their operating temperature range year-round. The compute system packs beneath the vehicle’s trunk floor, preserving cargo space for passengers. By co-designing custom silicon with the sensor and algorithms, Waymo achieves unmatched efficiency and performance. The latest system (right) processes high-fidelity data from 13 high-resolution cameras simultaneously and in real time, delivering exceptional low-light perception and revealing critical environmental details that traditional cameras (left) miss. WAYMO The blog authors explain that the company built a machine learning-primary architecture to run advanced neural networks at minimal latency. Then they plug in the best CPUs, GPUs, and accelerators to manage critical non-ML tasks like orchestration, data movement, and logging. While this new ASIC is in-house silicon, Waymo stresses its continuing partnerships with suppliers such as AMD, Micron, NVIDIA, Samsung, Sandisk, Socionext, and TSMC in its mission to develop the most capable autonomous computing system.