SV was never designed to be anything more than eyes-on, L2. ME is not claiming that the 30 TOPS will do anything more than L2. IMO, SV really just needs to be at the level of early FSD beta with an attentive human driver. Can it do that? Maybe.
I’m aware of that, I was pointing out specifically that SV1.0 has a very low ceiling limit due to its 30tops and the first version of SV that might show up in the US might be 1.0. So it’s almost certainly going to be inferior to FSD Beta.
And I see no reason ME could not decide to do SV 2.0 or 3.0 with more TOPS by adding extra eyeQ5 or eyeQ6 chips. So yes, SV 1.0 is currently limited to 30 TOPS but that does not mean that SV will always be limited to 30 TOPS. Heck, when the eyeQ7H chip comes out, ME could do SV with just 1 eyeQ7H and double the TOPS of the current SV!
Yes there’s already a SV 2.0 coming in VW/Porsche cars using 2x EyeQ6 (68 Tops apparently)
Chauffeur is designed to be the eyes-off system and it will have a lot more TOPS and more sensors than SV. Chauffeur will use 3-4 eyeQ6H according to their marketing material. If true that would be a total TOPS of 102-136, way more than the 30 TOPS of SV 1.0. I believe that is in the same ballpark of Tesla's HW3 computer. Now, I am not saying that Chauffeur (CH) will work. Maybe CH will totally suck. But the fact is that CH will have way more TOPS than SV. It is not hard limited to 30 TOPS. And if Tesla can do V12 end-to-end NN on HW3 with about 130 TOPS, then surely ME could do some more advanced NN/ML with 136 TOPS as well. We will see.
My main issue is that Mobileye isn’t focusing on SOTA algorithms or NN architectures, but rather legacy models including CNNs, etc.
We know that the only proof point we have of a generalized L4 cars are using very high compute (hundreds of TOPs if not up to or over 1k) and latest and greatest NN.
Especially heavy prediction network and ML based driving policy. These are very compute intensive and are the only method that has worked.
The issue is that mobileye WONT do an end to end model, or heavy ML in prediction and driving policy. That’s is the problem. This is exactly why they have been unable to produce a L4 car even though they started around when Cruise started.
Also, in that interview, Shai mentions that Chauffeur will be valuable because it will give people eyes-off in highway traffic jams. This strongly suggests that the initial version of Chauffeur will be a low speed traffic jam, L3 system.
I’m glad you caught that, the back pedaling has already started. Remember they were against traffic jams and saying it wasn’t useful in their marketing presentations.
So no, Chauffeur won't be L4 or eyes off everywhere. I have no illusions there.
Yet this is what they promoted 2 years ago.
Check out our #CES2022 virtual set for an exciting show! See how Mobileye is speeding toward a future in which the benefits of #autonomous driving will...
www.facebook.com
Can CH with ~100-130 TOPS, cameras, radar and lidar, do traffic jam L3? Probably.
I am very skeptical that ME can achieve the kind of MTBF numbers that they are talking about in their marketing material. Yes, SV on city streets will likely be delayed and yes CH will likely start as a neutered L3. I am certainly not suggesting that ME will solve L4 or have eyes-off everywhere, any time soon. I agree with you that ME will need to leverage the more recent advances in NN/ML. But I don't agree that they have a hard limit on their compute. The NIO ET7 uses the Orin chip and it boasts 1000 TOPS. There are other Chinese EVs that use Orin and have TOPS in the 500-1000 range. So there are OEMs using more compute in their EVs. So I doubt OEMs are placing a hard limit of 30 TOPS on ME. And ME is not placing a hard limit of 30 TOPS on themselves since they are already plans to use way more TOPS for Chauffeur and Drive and will likely use more TOPS when they need to. Lastly, ME has the eyeQ7H chip with 67 TOPS to start production in 2027 according to the marketing material. If that happens, ME could put 3-4 eyeQ7H chips in future vehicles and easily do more than 136 TOPS. So I don't think there is a hard limit to compute that will prevent ME from leveraging more advanced NN/ML. If ME fails to leverage more advanced NN/ML, it will be an intentional design choice on their part, not any hard limit that I can see.
Mobileye places a hard limit on themselves because their entire business model is to sell chips for literally acouple dollars(I believe eyeq3 was like around ~$15 and eyeq4 around ~$40). Nvidia on the other hand sells expensive chips for acouple hundred dollars. The 4x Orin is probably costing NIO around $2,000 give or take.
Mobileye can’t sell huge chips for that much, they are basically trapped by their own success. Because most of their customers just want basic adas.
This is why EyeQ5 went from 24 tops to 15 tops. By making small chips and different variants (low, mid and high). It would meet the needs of all their OEM customers. Think about it, the only one using a EyeQ5H chip right now is Zeekr and more recently Polestar 4 in china. Every one else EyeQ5 low or mid.
And since Mobileye decided not to go for huge powerful chips like nvidia did, they lost all the potential customers that want a huge powerful chip to run custom latest and greatest AI to Nvidia and Qualcomm.
But again my biggest problem with Mobileye. Take a look at their recent CES presentations, nothing new, no use of new AI systems and just a bunch of back pedaling. The thing about back pedaling is that it makes you play catchup. Look at Tesla and how they started focusing on simulation late. Same thing happened with Mobileye, did you see their simulation? It looks so primitive compared to other L4 sdc companies. This is what I mean by there’s just no way.
They too were against simulation at one point. Now they are on the simulation train, but they are so behind, others have moved to simulation 2.0 and 3.0. While they are trying to scrap together a simulation 1.0 video game graphics system.
One day they will realize that prediction network is needed and driving policy needs to be 50-75%+ ML. But again once they do they will yet again be playing catchup on that front.
www.bilibili.com
