Sabareesh Ravikumar

AI Compute @Waymo | Previously at  | Views are my own.

[Sabareesh Ravikumar @RSabareesh]

@rzidane360 About time! Forwarded to the team as required reading.

[Sabareesh Ravikumar @RSabareesh]

Anyone know a cracked engineer-in-the-making looking to intern on AI writing kernels for Waymo’s custom inference accelerator introduced at hot chips?

Lots of new problems, unlimited token budgets. Proof of work more important than credentials.

[Sabareesh Ravikumar @RSabareesh]

This is the output from our custom high speed lidar proceeding chips - before we do any ML. Beyond range, we track intensity, elongation and other metrics that give us an image like this - night or day.

[Daylen Yang @daylenyang]

@yunta_tsai @Bastien_Theron @wholemars actual lidar image from waymo 6th gen suite, no semseg

[Sabareesh Ravikumar @RSabareesh]

Good summary of our talk today @hotchipsorg from @ServeTheHome .

[STH @ServeTheHome]

We learned more about the Waymo Sensor Fusion processor at Hot Chips 2026 https://t.co/Fo7Zvs2qJo

[Sabareesh Ravikumar @RSabareesh]

@ServeTheHome Thanks for covering our talk!

[Sabareesh Ravikumar @RSabareesh]

@bookwormengr Thanks for attending our talk and glad you liked it!

[Sabareesh Ravikumar @RSabareesh]

@firstadopter Thanks for covering us!

[Sabareesh Ravikumar @RSabareesh]

@insane_analyst Excited to make it to the very interesting list. See you there!

[Sabareesh Ravikumar @RSabareesh]

After almost a decade at Waymo, I’m thrilled to publicly share a small glimpse of our work. Our sensor fusion chip makes real time decisions possible for the Waymo driver.

More technical details in my talk @hotchipsorg on Monday.

[Waymo @Waymo]

We’re putting our chips on the table. Learn more about our approach to engineering compute for fully autonomous driving and how we collaborate with industry leaders including AMD, Micron, NVIDIA, Samsung, Sandisk, Socionext, and TSMC to build the most capable computing system on the road: https://t.co/GlZZhBrw3F

[Sabareesh Ravikumar @RSabareesh]

@JeffDean Thanks for everything, Jeff. Expect a lot of CLs your way tomorrow as people try to get to dean number 1, for the last time.

[Sabareesh Ravikumar @RSabareesh]

Keynote coming from the boss this year. We’re also doing a technical talk in the automotive session later. See you there 👀

[Hot Chips @hotchipsorg]

Ten years after his Hot Chips keynote on silicon in self-driving cars, Daniel Rosenband returns with a new look at the road ahead.

Don’t miss his Waymo keynote:
“Compute in Motion: Challenges of Autonomous Driving.”

[Sabareesh Ravikumar @RSabareesh]

@0xBADB01E Sadly no one makes LP6 PHYs at N6. N6 is ironically a better choice for shoreline vs N3 as you’ll need at least 0.5x reticle for what you want or chiplets.

[Sabareesh Ravikumar @RSabareesh]

@itsclivetime Come for jalapeño, stay for Waymo 👀

[Sabareesh Ravikumar @RSabareesh]

@rzidane360 Too slow, let’s double the PE count. Oh wait.

[Sabareesh Ravikumar @RSabareesh]

@bubbleboi All fun and games till they try to design their own PHYs.

[Sabareesh Ravikumar @RSabareesh]

@jackcookjack Cool work! The paper shows how a hardware dot product unit can be designed for quantized values but the output quantization can take a non-trivial number of cycles - determine FP4 vs I4, doing bit manipulation for the signaling etc. What’s the overhead?

[Sabareesh Ravikumar @RSabareesh]

@kyunbit It has been crazy to follow along. Incredible team.

[Sabareesh Ravikumar @RSabareesh]

@andrewdfeldman Why disaggregate all of decode vs just the FFNs which avoids kv cache transfer?

[Sabareesh Ravikumar @RSabareesh]

@samwhoo Really enjoyed the blog and visuals. For the 4b version, I assume this isn’t doing any block quantization? Is it per tensor or per channel?

[Sabareesh Ravikumar @RSabareesh]

@AnasYMalas @ptrschmdtnlsn Ah nice. Will check it out!