1:12 “full self driving car hardware cost below 5000$”.
N.B. He is talking about real full self driving.
Thanks for the link.
My god, Amnon the Mobileye CEO sure knows how to spin small progress into sounding like they own the world. You have to listen very carefully to what he actually says. Throughout, he continually says things like "eyes off" driving, which sounds great, but then he quickly slides in caveats like only on "freeway on ramp to off ramp", which is what Tesla has had for six years at least? His revenues are expected aggregate revenues over eight years, not yearly revenues, but that isn't made very clear.
Having said that, it is interesting to hear what he has to say in detail. He makes good points here and there.
There are basically only two big players in this industry right now. Tesla and Mobileye. When you drive pretty much any other car brand and you get lane departure warnings or blind spot beeps, that's Mobileye. Most of their products do simple ADAS.
Mobileye's
target for freeway driving is max 80 MPH (Tesla's max is
currently 88 MPH).
Mobileye requires HD maps for FSD. They are also extremely risk adverse with an eye to lawsuit protection. They add stuff, seemingly to ensure lawsuit protection rather than to actually do anything too useful (that's my take). So they recommend adding front lidar and surround radars, and eventually surround
imaging radars (interesting distinction) just for redundancy when they get to FSD. Tesla has none of that, and Tesla FSD will just, presumably, stop working if a camera goes offline.
Mobileye needs 3 of their latest high performance chips for highway on ramp/off ramp. 4 chips for urban self driving.
When the hardware slides got shown (at 46m), I was surprised at how power hungry these system level boards are. The three chip board uses 275W, liquid cooled, and the 4 chip board (future product), 350W. Gone are the days I guess when Mobileye boasted about their 10W power consumption chips. It would be interesting to see what Tesla's board consume in power (only estimate I found was 72 watts from 2021, not sure how much of the entire board that included).
They continue to be way behind Tesla in getting hands free into the world. They only will have 150K cars with just highway hands free driving in 2023 (human still required to take over, so what Tesla had six years ago). They expect to have 1.2M in 2026. Their next step is to enable eyes off highway only driving where the human never has to take over, and that'll be rolled out for the first time in 2026. For that one, they have code where the car will pull over at the side of the road should anything bad happen or the human doesn't take over when requested (like when it leaves the freeway). So that means something like FSD is beyond 2026.
Both Tesla and Mobileye have their installed fleet send data back to the mothership for future training. But they do it very differently. Tesla has code running in each car such that when the car sees something it didn't expect, or a special campaign is run to gather certain data (like operating in driving rain), then high definition video is stored and sent back when the car is parked and connected to WiFi. Mobileye's OEM fleet isn't usually connected to WiFi, so they must rely on cellular, which is expensive, so they only send back compressed landmark information to make high definition maps (an HD map is just GPS along with dead reckoning for locations of things like signs, lanes, intersections). The two data gathering systems are very, very different. I honestly don't know how Mobileye trains their vision system since they don't have nearly anything as robust as Tesla's data loop by several orders of magnitude.
The presenter even denigrates Tesla's approach as saying that sending back high definition video would have increased cost so much it would have prevented gathering "at scale". Yeah, sorry, Tesla also has a well thought out system, which mitigates cost and/or spends the money to do it right.
"You cannot have an eyes off system without high definition maps". (54m in video). Ummm, Tesla (and I) disagree.
Someone upthread said Tesla vs Mobileye is like Apple IOS vs Android, or MacOS vs Windows. There are certainly some similarities. Reading between the lines, Mobileye has had trouble convincing OEMs to implement the full autonomous vision system, and if their target cost is $5,000 (that includes all hardware, sensors and Mobileye chips), I can understand why. That's a lot of cost to add to a car. No one will do it unless they are forced to for competitive reasons. Tesla's FSD isn't fully realized enough to force such adoption. Tesla highway AP is good enough to be a competitive factor, but I'm not sure enough people know about it and/or if the OEMs know their customers are leaving them because of it. There's a lot of denial in the industry.
So Tesla has an advantage over Mobileye in that their actual incremental costs per car is lower than Mobileye's $5000 (cameras are cheap and they own their own AI chips), probably even if you amortize the huge costs of Dojo and the A100 GPU clusters, the AI team, etc. Moreover, Tesla is now spreading the cost of most of that AI investment into another completely different, but no less huge market, humanoid robots. Did you know that half of the human brain is used for vision processing? Turns out that Tesla has attacked one of the hardest AI problems (vision perception) first.
Normally, Mobileye would have a cost advantage because their R&D is spread over a much large number of OEM cars and manufacturers. BUT there are costs dealing with all the heterogeneity. Every unique car, every manufacturer requires hand holding. You need a sales force to sell Mobileye versus the competition (there are other competitors out there). And you don't have systems level integration opportunities that Tesla has. And even worse for Mobileye, Tesla is increasing their vehicle base exponentially. Tesla is years ahead in installed based for both highway driving and FSD now, and I don't see their lead diminishing, if anything it'll get even bigger over time.
And you can kind of see this in the Mobileye presentation. Their simulation environment to train their AI is rudimentary compared to Tesla's. They don't have a tight feedback loop. It just looks to me that Tesla is ahead now and pulling away rapidly.