Tesla FSD 12.X and Beyond Will Achieve Faster Rate of Improvement

by

Here is a video discussion of the most important point about Tesla FSD 12.X.

The indications are that the rate of improvement for the Tesla FSD 12.X all neural net system will greatly increase.

There was over 300,000 lines of hard coding removed from Tesla FSD.

It was reported that Tesla had purchased 15,000 Nvidia H100’s in 2023. H100s are 11 times as powerful for AI as Nvidia A100. This means Tesla has the equivalent of at least 165,000 A100s.

“The [New York] governor is correct that this is a Dojo Supercomputer, but $500M, while obviously a large sum of money, is only equivalent to a 10k H100 system from Nvidia,” Musk wrote in the post on X. “Tesla will spend more than that on Nvidia hardware this year. The table stakes for being competitive in AI are at least several billion dollars per year at this point.”

It appears that Tesla will likely have 50,000 to 200,000 Nvidia H100 equivalents by the end of 2024 and 10,000 of that will be Dojo training. Tesla spending $3-5 billion would be about 100,000 to 200,000 Nvidia H100s. This would be 1 -2 million Nvidia A100 equivalents.

This means that Tesla’s training compute should be well over half of the 100 Exaflop goal that Tesla described for October, 2024. It should take some months to install but Tesla has had 4-6 months to install the Nvidia H100s.

The Tesla FSD training should be ten times faster than in early 2023.

Tesla is adding about 100 million miles per month of FSD training data in 2024. In 2022, the estimate was in the range of about 6 billion miles of training data.

If all 500,000 FSD beta users were always using FSD all the time they would each add about 1000 miles per month. This would be 500 million miles per month of data. Tesla FSD giving a better experience will increase adoption, sales and usage. Tesla is starting to sell FSD in China and will be selling in Europe later in 2024. Tesla will add about 2.2-2.5 million cars to the 5.5 million car fleet. This will increase the FSD sales and adoption.

Whole Mars points out 400,000 – 500,000 users out of 2 million Teslas in North America, 20% – 25% of users who can run FSD are running FSD. That is an astonishing penetration rate given that this is an expensive $12,000 or $2,400 a year option today. Can we double that to 40% – 50% of users who can run it? With FSD Beta 12 I [Whole Mars] think It could happen this year.

Whole Mars is driving FSD 12.X now.

11 thoughts on “Tesla FSD 12.X and Beyond Will Achieve Faster Rate of Improvement”

  1. [ What’s the computing constraint for FSD capability for Tesla? (is it training data (non ordinary situations, quality or amount (?)) from miles driven through FSD users or is AI training hardware on its top utilization levels (compared with optimized technical progress, neglecting availability or financial limits)?

    ~500.000 FSD users are contributing ~$1.2Billion/year, that’s ~30.000 Nvidia H100(&cars hardware, &software development, &data center installation (?)), ~100ExaFlops/yr_500kUsers, ~~5-6Billion data miles/yr_500kUsers

    (thx) ]

  2. I absolutely love driving, though sometimes I’d like the option of FSD. I like going for long drives on my own; it’s meditative. Travel restrictions are for dictatorships.

    With that said, it seems FSD technologies will have utility beyond the auto industry.

  4. The average person drives for about 10 years longer than they really should just because taking someone’s car privileges away based on suspicion that they may be past it is a big step. When my father went beyond driving safely my mom had to put her foot down and told him she was doing the driving from then on. It was a bit easier when I took over for her but not much. I like the independence of having a car and I don’t mind driving in places I’m familiar with but honestly I’d rather be doing something else while being chauffeured, especially in places I’m not familiar with. There are a LOT of psychotic people out there and I’d rather a machine deal with them.

    I’m not old yet but I’m getting older and when I get past driving it would be nice to have a means of transport that doesn’t involve being crammed into a bus with weirdos or needing to sell a kidney each time I want to pay a taxi driver. I’m old enough to remember when taxis were actually sort of affordable.

  5. I rarely drink, I’m a guy, and white as the snow, and I’d love a FSD car.
    I enjoy driving about 5% of the time, the rest, I’d rather be taking a nap, or watching Youtube/tiktok.
    Plus many people can’t drive, My dad is 68, and his vision isn’t great anymore, so I’m hoping FSD gets cracked soon (I’m currently trying to convince him to buy a Comma 3x).

  6. I haven’t had an accident in over thirty years. Why would I trust FSD? Why would I need FSD? I enjoy driving. FSD is for drunks, women, and Asians.

Comments are closed.