Elon Musk Claims on Tesla Chips Don't Match the Reality

>Ask yourself why nVidia would feel the need to put out anything given that they and Tesla are no longer collaborators.

nVidia must be short-selling Tesla stock. If Hacker News has taught me anything, it's that we can rest assured about our TSLA investments because the only reason anyone ever says something negative about - or disagrees with - Musk or Tesla is that they're short sellers and we've just gotta ride it out.

Except if you screw up and cut the cable too short it's going to really hurt your finances, and maybe more importantly, your time to market. There's quite a lot of risk involved but I guess a big enough budget can help mitigate some of that.

Practically speaking there is no power limit. The traction motors on Tesla's cars themselves will pull two orders of magnitude more power than these chips will.

Granted, there are other factors (obviously), but it seems odd to be quibbling about even 400W on a car with a 100KWh battery pack. 0.4% difference in consumption accounts for less than 2mi of range.

I'm watching the video. Highly interesting.

That level of power use would take *noticeable" dents into the range of the car, especially in inner city driving.

Furthermore that's a lot of thermals to get rid of. 500w, crikey, speaking as someone who water cools his PCs, forget about retrofitting older cars! (That was one of the criteria!)

The existing units sit behind the glove box in front of the firewall, and are rather compact. The thermal budget was a hard one.

Nvidia are full of crap on this one.

A bit apples and oranges, though. The ASCI Red is a general-purpose computer that ran whatever you wanted via the full x86 instruction set. The Xavier and Tesla chips are basically very specific FMADD modules that can't do much else.

I didn’t watch the whole livestream, but it could be one talking about just the specific chips versus the other the whole system?

They're in the right ball park to see wins.

Samsung fabs it. Do you mean chip design and tape out NRE?

Tesla spent several years designing their current chip, finishing about 1.5 years ago. They hired a industry guru to help them do it, I’m sure the financial investment was massive. Elon stated somewhat wryly “their entire expense sheet is FSD” which is to say their R&D spending in FSD dwarfs their other R&D spending.

They are also already working on a Gen2 of their just released chip for the last year plus.

For Tesla the hardware design is essential to achieving their goal. If they get FSD in the real world before everyone else, they can print money with the Tesla Network. And they simply can’t get there first using general purpose off the shelf hardware.

It’s a core part of the competitive advantage of the Tesla fleet. The fact that you can slot this new hardware into any AP2.5+ car shipped in the last couple years is an amazing perk of owning a Tesla. No other carmaker comes close to offering this kind of long-term support with software and hardware updates to a late-model-year car.

I think you're massively under-estimating the difference in specialization. Tesla know the resolution of the data going into the Neural Nets, they know the depth of the data, they know the types of non-convolution stages (pooling, ReLU Sigmoid etc). They've probably got a fixed number of layers too. So they're basically completely getting rid of the control logic. That's an enormous chunk of the problem, and lets' remember these are peak performance numbers - the Tesla programming model may be a disaster. Hell, they could probably have bought the IP for the actual computation core.