5 ms·

It’s a chip for doing exactly the convolutions they want to do a whole lot of ... Nvidia gpus present a very different and more flexible programming model — why

It’s a chip for doing exactly the convolutions they want to do a whole lot of ... Nvidia gpus present a very different and more flexible programming model — why is it not the expected case that Tesla’s hardware would have the price and wattage advantage for the problem they intend to throw at it ...?

kingosticks 7y ago
Do they have the scale to achieve any sort of price advantage? 14nm NRE costs are huge.
- monocasa 7y ago
  They're in the right ball park to see wins.
- zaroth 7y ago
  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.
  - madengr 7y ago
    I assume it will be a free upgrade to those who paid for the self-driving hardware?
zaroth 7y ago
Absolutely it is possible to do better by specializing, that doesn’t mean we generally expect an auto-maker to be able to successfully make nVidia’s best effort in the same marketplace obsolete in their first go at designing their own chip. This isn’t quite like GPU vs ASIC for Bitcoin mining, for instance, because nVidia Xavier is already a special purpose chip for doing the Int8 calculations. Tesla isn’t competing against a graphics card being repurposed for FSD, they are competing against nVidia’s FSD offering, and using an industry standard metric (Int8 TOPS) to measure their performance.
- Traster 7y ago
  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.

kingosticks 7y ago

Do they have the scale to achieve any sort of price advantage? 14nm NRE costs are huge.

monocasa 7y ago
They're in the right ball park to see wins.
zaroth 7y ago
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.
- madengr 7y ago
  I assume it will be a free upgrade to those who paid for the self-driving hardware?

zaroth 7y ago

Absolutely it is possible to do better by specializing, that doesn’t mean we generally expect an auto-maker to be able to successfully make nVidia’s best effort in the same marketplace obsolete in their first go at designing their own chip. This isn’t quite like GPU vs ASIC for Bitcoin mining, for instance, because nVidia Xavier is already a special purpose chip for doing the Int8 calculations. Tesla isn’t competing against a graphics card being repurposed for FSD, they are competing against nVidia’s FSD offering, and using an industry standard metric (Int8 TOPS) to measure their performance.

Traster 7y ago
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.