-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Waymo
- Thread starter Daniel in SD
- Start date
For training the most probably use GCP and TPU v5:s. Google Cloud probably has more GPU power than everyone else in the S&P 500 combined.So.. long thread, but was pointed here when I asked these questions:
Any pointers to specific post/answers appreciated, as reading 189 pages might take a while
Google announces the Cloud TPU v5p, its most powerful AI accelerator yet | TechCrunch
Google today announced the launch of its new Gemini large language model (LLM) and with that, the company also launched its new Cloud TPU v5p, an updated
techcrunch.com
For training the most probably use GCP and TPU v5:s. Google Cloud probably has more GPU power than everyone else in the S&P 500 combined.
Google announces the Cloud TPU v5p, its most powerful AI accelerator yet | TechCrunch
Google today announced the launch of its new Gemini large language model (LLM) and with that, the company also launched its new Cloud TPU v5p, an updated
Sure, but what do they dedicate to it for a training session?... that infrastructure is shared amongst paying customers...
diplomat33
Average guy who loves autonomous vehicles
So.. long thread, but was pointed here when I asked these questions:
Any pointers to specific post/answers appreciated, as reading 189 pages might take a while
Waymo is very tight lipped on their compute. So I don't know if there is a lot of specific info out there. We do know the cars have a custom in-house computer.
This is what @Bladerskb shared in another thread about their on-board compute in the cars:
"7 years ago they were using 100 TOPs. Today they are definitely in the hundreds of TOPs and maybe up to 1,000 TOPs. But I don't see them being much over 1k. But no one knows for-sure."
I don't think anyone outside Waymo really knows the answer to that. AFAIK, Waymo has not shared that info.
Let's just assume that Waymo and other internal Alphabet teams like Google AI/DeepMind, search and Youtube don't have to worry about compute budgets or compute availability... Google has been a leader in AI för 10+ years and has leading custom silicon and software that scales and they make a lot of money. Google net profit is almost at par with Tesla's revenue.
Not suggesting that they are necessarily constrained (well they are, because compute not infinite, and they do have paying customers), just wondering if we had a ballpark of what was thrown at it... but sounds like nobody outside Waymo really knows.
Thanks.
diplomat33
Average guy who loves autonomous vehicles
Waymo is going to Washington DC for data collection and testing, to further improve the Waymo Driver:
Last edited:
DanCar
Active Member
As you know, Waymo is constantly updating their hardware including their compute. They have custom Google edge tpu hardware on board.
For training they have vast Google datacenters available to them. I'll guestimate a billion dollar per year datacenter budget. As other's have said, Google's TPU and GPU resources are vast and growing rapidly.
I can guess that Waymo's onboard processing power is >10x of a Tesla.
Are they actually having cars drive cross country to DC? Like, a real road trip?
That would be VERY interesting to see.
jabloomf1230
Minister of Silly Walks
That's a wonderful guess. Google probably has orders of magnitude larger computing power than Tesla. But how much of that power the tiny Waymo unit has access to is unknown. For example, Alexa could have access to the enormous compute resources that Amazon has. But based on Alexa's half-witted behavior, I tend to doubt it.
DanCar
Active Member
Waymo is/was Larry and Sergeys pet project. That should give a good clue how much access to compute resources they have .
EVNow
Well-Known Member
These things are difficult to tell - esp. because compute not given to Waymo can be sold and generate revenue. When I was in MSFT, it was difficult to get capacity internally. They prioritized customers.
DanCar
Active Member
When I was at Google, Waymo didn't have a shortage for most years. There isn't much of shortage at Google for most things, except TPU. They were told to cut back on the log file size planned increases at some point.
Is Waymo in the process of expanding their Phoenix valley territory? I saw two Waymo's driving surface streets a bit north of the Scottsdale airport. That's a few miles north of their mapped territory.
diplomat33
Average guy who loves autonomous vehicles
EVNow
Well-Known Member
I know Google has had difficulty growing the cloud business - at Azure when I was there the demand was insane.
diplomat33
Average guy who loves autonomous vehicles
Yeah, she is always great in these types of chats. She focuses on the operations side and makes it easy to understand. I also really Dolgov's interviews when he does them because he gets into the technical side more. I am very interested in the technical part of vehicle autonomy so I love to learn about that stuff.
Did you notice her reference to Tesla FSD being a sort of L2/L3 hybrid. She says that Tesla FSD is confusing because it is not always clear who is the "driver". She is not wrong. When FSD is on, it can feel like FSD is the "driver" since it is doing all the steering and braking. But at the same time, the human still needs to supervise so the human is actually still the "driver" and when there is an intervention, it can quickly change as the human has to take over steering and braking again. And that is why Waymo focuses on L4. In a Waymo, the car is always the driver and the human is always the passenger. So there is never any confusion about who the driver is.
Well, she's wrong (or generous) that FSD would be L2/3. It's L2. Period. L3 includes OEM liability.
I liked the answer at 08:30 the best. And the second was the stuff about "we could open in 20 cities, but that not how we do it" (paraphrasing) towards the end.
Edit: Link - Every major city
She also claimed they couldn't deploy the Zeekr now, for legal reasons (NHTSA I assume) still.
Last edited:
diplomat33
Average guy who loves autonomous vehicles
I think she was being generous since there are certain ODD now where V12 is arguably getting close to L3 capability.
J3016 actually says nothing about liability. So technically, L3 does not include OEM liability. Some argue that OEM liability is implied in L3 because the system is performing the entire DDT. But I think for L3, liability is tricky since the L3 is responsible for the DDT but the human is responsible for the DDT-fallback. When the L3 is engaged, it is performing the entire DDT so logically the OEM would be responsible during that phase but what about when the L3 asks the human to take over to perform the DDT-fallback? If the human fails to take over, they would be liable or is the OEM liable because there was a failure in the driver monitoring system that did not adequately hand over the task to the human? We can see how liability during the transition from L3 to the human could be unclear. For L4/L5, I think OEM liability is much more cut and dry since the system performs both the entire DDT and the DDT-fallback. Again, OEM liability is not mention in J3016 for L4/L5 either but I think it is strongly implied since the system is always responsible for the driving tasks.
I think her answer is basically the "quality over quantity" argument. Waymo could do 20 cities (quantity) if their goal was just to claim a "first". But then the quality might suffer as each service might not be very big or as good as it could be. Waymo prefers quality over quantity because Waymo wants to make each service area really good and meaningful in the community before moving on to the next city. They don't want to just spam service areas and have each service area be tiny and not very useful. I appreciate her answer. I like quality over quantity. My only concern is that it could cause Waymo to scale too slow if they chase perfection too much in each service area. And when the Waymo Driver is generalized enough and safe enough, I hope to see Waymo scale faster. I don't want Waymo to just focus on perfection in a few cities at the expense of scaling to other places.
Yeah, I think NHTSA has rules about deploying a car with no driving controls. I believe the manufacturer has to get a special dispensation form NHTSA to deploy a vehicle with no driving controls above a certain speed. Waymo could get around this by simply deploying a variant of the Zeekr with driving controls. Then Waymo could still deploy it driverless like they are doing with the I-Paces, As long as the vehicle has a steering wheel and pedals, Waymo would be ok.
Similar threads
- Article
