[nine_k]

Success rates may differ.

Do you know that AMD makes powerful GPUs that even have a general purpose programming interface called ROCm? Now try to persuade ML practitioners to switch to that ftom NVidia GPUs and CUDA.

[Aerroon]

And what would happen if Nvidia got banned? ROCm would take over. All the development resources would be going into that.

That's probably what's going to happen in China.

[nine_k]

ROCm likely won't take over, because of its flaws. Eventually, of course, either it or something else, would emerge as a solution.

But it will take time. Buying time is what this ban is trying to achieve. Slowing down the rivals.

[londons_explore]

I think you overestimate how long such a change would take.

If Nvidia was banned tomorrow, I think within 6 months another leader would emerge and be supported by all the big libraries.

[IshKebab]

ROCm is already supported by Pytorch apparently.

[londons_explore]

My understanding is that while it is theoretically supported, the vast majority of random pytorch projects you find on GitHub will need tweaks before they'll run in ROCm

[latchkey]

https://pytorch.org/blog/experience-power-pytorch-2.0/