I'm going through the saved video.
The early part is pretty boring. Too much detail on chip design basics which I already know. Feels like filler to impress people who don't know the subject. It's not impressive at all, by the way. Many minutes of this stuff add up (for those who know the meaning of the details) to "Yes, we designed our own custom chip".
Looking at the investor thread, a lot of people were saying "Does anyone understand this?" Well, I do, it's not complicated, and this is really a pretty basic tutorial. (So, no, they aren't revealing any secrets.) The basic math here is not hard.
(It's math I don't *like* -- mathematicians have tastes and this isn't my flavor of choice -- but it's easy math conceptually, so I followed 100% of it. And saying that it's easy is not saying that optimizing it is easy -- optimizing it is fiendishly tricky bit-twiddling, which is why I don't like it, and that's what Bannon & company get the big money for. But it is 99% linear alegbra.)
HARDWARE:
It does seem that they're the only company with suitable hardware. ("You only have to run faster than the bear.") In addition, they should probably sell their Neural Network Chips, since apparently (according to the first speaker, Pete Bannon), they're literally the only company making one, and there are an awful lot of applications where they'll be extremely attractive. That's a potential real source of money if they bother to sell it to others.
As expected, they detect internal hardware failures by feedback from sensors.
As expected, they've attempted spacecraft-style redundancy (easy to attempt, who knows whether they succeeded).
"Advanced tone mapping" and "Advanced noise reduction" have no details (and here details would have been interesting, since you can lose data in stuff like this).
The cameras all go straight to the chip (only), and the video encoder feeds out any video which some other car component might need (like the backup camera display).
--- It appears that they have made *multiple* neural networks for specific purposes? The neural network accelerator can run one neural network and then switch to another? I don't think that's used in live driving?
All their multiply operations are 8-bit -- that's an interesting choice -- done for energy reduction apparently. They think this is enough accuracy. That's a daring decision, and they might be wrong. They might be right, though. I'd be interested to see the evidence behind this.
Got rid of most of the bookkeeping waste from the classic Von Neumann design (as they should)
Solid neural network compiler, makes sense
-- Yes, there seem to be multiple neural network programs resident in SRAM. The computer can be told "Run this neural network on this photo", and gives a result. Then switch to the next processing job with a different neural network.
They keep using the "narrow camera neural network" as an example. Apparently there's one neural network for each camera, which takes the image and converts it into some sort of abstract data result. Then some other neural network must run on those data results?
Silicon cost is 80% lower than other chips. Speed is 21x NVIDIA. (They should sell these chips and boards.) Power requirements are at most slightly higher.
-----
Yes, this is the best neural network computer in the world, by far. It will be quite hard for anyone to duplicate this quickly. It's potentially a cash cow if they decide to sell the chips, but they are apparently not doing so.
Musk says "All you need to do is improve the software", which is like saying "All you need to do is the hard part" or "We did everything except self-driving". Funny. I'll have a separate comment when they switch to the next presenter.
First question was good, and the answer was good -- the NN accelerators can be used for a different type of problem.
Musk later says that it's just specialized for self-driving -- explaining that NVIDIA has to be more "general purpose" which supposedly forces their chips to be less efficient. I don't really believe that. The 8-bit restriction may be serious in some applications, but it sounds like the system is actually quite general-purpose. I don't see any reason why it couldn't handle a different NN application as long as it was heavy on convolution/deconvolution and didn't require longer-than-8-bit multiplication.
(IMO, a two-year serious NN chip program could definitely make a similar or better chip, but NVIDIA or Intel or someone would have to actually do it. Musk says three years, but Musk will have a better system in three years. IMO, and this is my guess, the chip companies don't have the sales volume for it to be worth doing the development -- the "million cars a year" is what makes it commercially sound to finance it.)
Musk points out that power reduction matters for car range. Says most of the robotaxi market will be in cities (well, that's right).
Fabbed in Austin Texas by Samsung.
Question about whether the weights could be encrypted/protected from just being lifted out of the car. Basically, not possible. But it should be hard to crack. Anyway, Tesla can constantly replace its weight numbers because it has the fleet, everyone else would have to steal it repeatedly
Musk points out that simulated driving is ridiculous because the goal is to identify edge cases. "The real world is weird and messy".
Good.
