Waymo details 5nm sensor-processing ASIC with over 1,000 TOPS for robotaxi perception
- Waymo says its purpose-built 5nm ASIC delivers over 1,000 TOPS for front-end sensor processing and machine-learning models, processing raw lidar, radar, and camera data before it reaches the core ML system.
- The chip runs specialized accelerators for sensor fusion, including temporal denoising for low-light perception, and Waymo says its latest system processes data from 13 high-resolution cameras simultaneously in real time.
- Waymo says it has increased raw compute 20x in eight years and optimized the stack to reduce "pixels-to-actuation" latency for onboard driving decisions made within milliseconds.
- The compute system uses two independent, parallel-running units so one can take over after a fault, and connects to the vehicle's liquid-cooling system for vibration, shock, and temperature exposure.
- Waymo says its heterogeneous architecture combines custom ML silicon with CPUs, GPUs, and accelerators for orchestration, data movement, logging, and inference; listed partners include AMD, NVIDIA, Samsung, Socionext, and TSMC.
Hacker News opinions
I think Waymo is the impressive autonomous-driving company, even if it gets less media attention than rivals. If we get there, I would not be surprised if Waymo is why.
My first Waymo ride was mind-blowing, then it quickly became an ordinary car ride. I wish more people could try one and see that it already works.
I think Waymo intentionally avoids the spotlight. Uber and Cruise drew negative attention and folded, so under-promising and delivering makes sense to me.
Waymo looks far ahead in sensors, vehicles, training data, simulation, infrastructure, operations, and markets. I still wonder whether municipal approvals, training, or manufacturing are the real bottleneck, especially with lobbying slowing rollout in NYC, Chicago, and DC.
I do not see slower robotaxi approval as automatically unfortunate. NYC and Chicago have public transit that can move more people with less car dependence than a robotaxi fleet.
I had a good Waymo ride in San Jose until it dropped me across a 10-lane road from a hospital instead of turning into the property. That drop-off choice made sense in San Francisco but was awful in that suburban setting, while Lyft drove to the actual destination.
I see Waymos actively testing around Chicago now. They started doing that a few months ago.
I am skeptical Waymo will work well in NYC. It seems to need some cooperation from drivers and pedestrians, and I cannot picture New Yorkers giving it much.
I expected commodity hardware, so the custom chip is interesting. I would like to know what the chip and developer tooling are actually like, especially since proprietary AI accelerators such as Meta's MTIA lack an open backend.
This post does not say much beyond marketing claims. Tesla has given more detailed talks about its hardware and AI stack.
Autonomous cars look like unusually demanding edge computers to me: limited power and cooling, unreliable connectivity, and strict latency limits. Designing hardware around that workload makes more sense than forcing a general-purpose platform to handle everything.
The camera example does not look high resolution to me and appears noisy. If that represents the visual pipeline, I can see why Waymo still relies heavily on costly lidar.
I suspect the model trains on raw or near-raw images because image enhancement can introduce discontinuities. What looks better to a person may not be better input for the model.
I noticed Waymo says ML rather than AI throughout. Given the current baggage around the term AI, I can understand why a marketing team would choose ML.
I do not view a large custom-compute setup as a win by itself. Tesla's FSD computer uses less power, and I think that comparison matters when judging whether Waymo can scale.