Updated
Updated · Waymo · Aug 20
Waymo Unveils 5nm ASIC Delivering 1,000 TOPS for Autonomous Driving
Updated
Updated · Waymo · Aug 20

Waymo Unveils 5nm ASIC Delivering 1,000 TOPS for Autonomous Driving

3 articles · Updated · Waymo · Aug 20

Summary

  • Waymo said its new purpose-built 5nm ASIC handles front-end sensor processing and neural-network workloads in real time, delivering more than 1,000 TOPS for its autonomous driving stack.
  • The chip is designed to cut milliseconds from “pixels-to-actuation” latency by processing and fusing raw lidar, radar and camera data onboard, a requirement for driverless operation without human backup.
  • Waymo said its latest compute system now processes high-fidelity feeds from 13 high-resolution cameras simultaneously, with custom accelerators improving low-light perception through features including temporal denoising.
  • The ASIC sits inside a broader heterogeneous, redundant compute architecture that Waymo says has scaled raw compute 20-fold in eight years while preserving battery efficiency, trunk space and silent operation.
  • Waymo framed the silicon push as part of a decade-long co-design effort with partners including AMD, NVIDIA, Samsung and TSMC as it prepares for broader uses of the Waymo Driver.

Insights

Why did Waymo abandon standard off-the-shelf chips to secretly build its own autonomous brain?
How does Waymo's custom 5nm chip predict unpredictable human behavior on the road before it happens?
If dual redundant onboard systems fail, what ultimate fail-safe keeps Waymo's driverless robotaxis from crashing?