Waymo Unveils custom 5nm ASIC powering autonomous rides

Waymo recently published a blog post detailing its approach to onboard compute systems for fully autonomous vehicles, highlighting custom silicon and collaborations that enable level 4 self-driving without human oversight.

Key Takeaways

• Waymo engineered its compute stack for ultra-low latency responsiveness, scaling performance 20x over eight years while processing sensor data in milliseconds for real-time decisions.
• The system incorporates ruggedized hardware with liquid cooling and built-in redundancy featuring dual independent engines to maintain safety in extreme conditions without driver backup.
• A new 5nm ASIC delivers over 1,000 TOPS of ML performance, optimizing sensor fusion and neural network inference across lidar, radar, and camera inputs in a heterogeneous architecture.

Deep Dive into Waymo Compute Architecture

Waymo shifted from off-the-shelf components to a purpose-built system that balances ML accelerators with CPUs and GPUs for orchestration tasks. This design maximizes efficiency in low-batch regimes common to autonomous driving. The 5nm ASIC handles front-end processing including temporal denoising for low-light perception before feeding data to inference engines, according to the Waymo blog post.

Technical Requirements and Innovations

The three core requirements of responsiveness, ruggedization, and redundancy address the demands of unsupervised operation. Integration with vehicle cooling sustains peak performance across temperature extremes while preserving trunk space and minimizing noise for riders.

Business Impact and Opportunities

Companies in the autonomous vehicle sector can leverage similar custom silicon strategies to reduce latency and power consumption, opening monetization paths through robotaxi fleets and logistics partnerships. Implementation challenges such as high development costs for ASICs may be solved via industry collaborations with chipmakers, enabling scalable deployment in urban mobility markets. Regulatory considerations around safety redundancy align with emerging AV standards, while ethical implications emphasize proactive fault tolerance to build public trust.

Future Outlook

Industry shifts toward ML-primary architectures will accelerate as competitors adopt heterogeneous compute designs, potentially lowering per-mile operating costs and expanding commercial applications in delivery and ride-hailing by the end of the decade.

Frequently Asked Questions

What makes Waymo's compute different from driver-assist systems?

Waymo's system operates without human backup, requiring higher performance, full redundancy, and onboard real-time processing for complete driving tasks.

How does the 5nm ASIC improve performance?

The ASIC provides over 1,000 TOPS dedicated to sensor data fusion and neural network execution, optimizing bandwidth and enabling advanced models like transformers.

What are the main business opportunities from this technology?

Opportunities include scaling robotaxi services, forming silicon partnerships, and achieving cost efficiencies through optimized power and space usage in commercial fleets.