Waymo has designed a robocar chip to stay ahead of Tesla

Waymo, the Alphabet-backed robo-taxi startup, has developed a custom AI chip to help its autonomous vehicles react faster and more safely to their surroundings. The new ASIC, built on TSMC’s 5 nm process technology, is specifically designed to handle both traditional machine learning and modern transformer models, which are used in AI chatbots and image generation models. The chip is optimized for speed, reliability, and redundancy, taking into account over 200 million miles of autonomous driving data.

Waymo replaced Intel FPGAs, which are commonly used for sensor processing in low latency applications, with this custom silicon. FPGAs are difficult to program and lack the compute density of dedicated AI hardware. The chip can perform over 1,000 TOPS of AI performance, which is comparable to Nvidia’s Drive AGX Thor platform. The chip incorporates multiple layers of redundancy, including liquid cooling and a backup ASIC, to ensure reliability in extreme conditions.

While this chip handles AI tasks, non-ML functions like orchestration, data movement, and logging are handled by components provided by partners like AMD, Micron, Samsung, Sandisk, and Nvidia. Waymo plans to share more information about its ML accelerators during the upcoming Hot Chips conference at Stanford.