2 ms·

What is Intel's future? I find myself asking this lately. The desktop market is strong but niche, and their datacenter numbers are waning. Perhaps there's a fut

What is Intel's future? I find myself asking this lately. The desktop market is strong but niche, and their datacenter numbers are waning. Perhaps there's a future where an Intel/AMD consortium could take on the dominance of Nvidia in the GPU ML space. It really comes down to creating a viable alternative to CUDA, which maybe they could do together.

smoldesu 3y ago
I think Intel will be fine. Their far future is probably based around fabbing mid-range chips for other customers, which should be pretty lucrative (with or without Samsung as a competitor). Maybe they'll rekindle their RISC-V research once their ARM partnership winds down, or double down on their GPU line. They have a lot of options right now. > Perhaps there's a future where an Intel/AMD consortium could take on the dominance of Nvidia in the GPU ML space. The thing that is cementing Nvidia's dominance is industry hostility. If you could ship a truly cross-platform GPGPU driver to every modern device, CUDA would be a harder sell. Instead, the best you can get are intermediary libraries like ONNX that do their best to cherry-pick and expose common and accelerated operations. CUDA eats it's lunch in featureset and optimization. To fix this, you have to bring Microsoft and Apple to the table too. It is purely a matter of investment and cooperation, neither of which seems to exist right now.
- foota 3y ago
  Isn't Microsoft pretty invested in onnx?
  - smoldesu 3y ago
    They are. The problem isn't them, it's everyone else; their execution providers are all over the place, with many lacking fundamental features and others with insane compilation requirements. And when you compare it to CUDA... they're in different leagues, literally. ONNX is solely looking at AI acceleration, not the field of GPGPU compute. If you want synchronized memory pages or generic GPU compute APIs, ONNX is not even part of the discussion. It's a great effort, and one I'm definitely rooting for. I wouldn't expect it to dethrone Nvidia though, even if someone manages to beat their hardware.

smoldesu 3y ago

I think Intel will be fine. Their far future is probably based around fabbing mid-range chips for other customers, which should be pretty lucrative (with or without Samsung as a competitor). Maybe they'll rekindle their RISC-V research once their ARM partnership winds down, or double down on their GPU line. They have a lot of options right now. > Perhaps there's a future where an Intel/AMD consortium could take on the dominance of Nvidia in the GPU ML space. The thing that is cementing Nvidia's dominance is industry hostility. If you could ship a truly cross-platform GPGPU driver to every modern device, CUDA would be a harder sell. Instead, the best you can get are intermediary libraries like ONNX that do their best to cherry-pick and expose common and accelerated operations. CUDA eats it's lunch in featureset and optimization. To fix this, you have to bring Microsoft and Apple to the table too. It is purely a matter of investment and cooperation, neither of which seems to exist right now.

foota 3y ago
Isn't Microsoft pretty invested in onnx?
- smoldesu 3y ago
  They are. The problem isn't them, it's everyone else; their execution providers are all over the place, with many lacking fundamental features and others with insane compilation requirements. And when you compare it to CUDA... they're in different leagues, literally. ONNX is solely looking at AI acceleration, not the field of GPGPU compute. If you want synchronized memory pages or generic GPU compute APIs, ONNX is not even part of the discussion. It's a great effort, and one I'm definitely rooting for. I wouldn't expect it to dethrone Nvidia though, even if someone manages to beat their hardware.