Waymo has designed a robocar chip to stay ahead of Tesla

Waymo, the Alphabet-backed robo-taxi startup, has unveiled its first custom AI chip designed to help autonomous vehicles make split-second decisions on the road. In a blog post, Waymo revealed that the chip, built on TSMC’s 5 nm process technology, is optimized to run both traditional machine learning algorithms and modern transformer models.

The company claims the chip incorporates over 200 million miles worth of autonomous driving data and is engineered to maximize responsiveness, reliability, and redundancy. Prior to this, Waymo had used Intel FPGAs for sensor processing, but the custom ASIC offers a more efficient solution for real-time AI processing. Waymo's chip is capable of handling over 1,000 TOPS of AI performance, which is comparable to Nvidia's Drive AGX Thor platform.

The startup emphasizes the importance of minimizing latency, allowing the vehicle to process data within critical milliseconds. To ensure reliability, Waymo has implemented multiple layers of redundancy, including liquid cooling and dual ASICs per vehicle. While the custom chip handles AI tasks, other non-ML functions are handled by components from partners such as AMD, Micron, Samsung, Sandisk, and Nvidia.

Waymo will provide further details on its ML accelerators at the upcoming Hot Chips conference in Stanford.