> there is really nothing that "normal" AI requires that is bound to CUDA. pyTorch and Tensorflow are backend agnostic (ideally...).
There are a lot of optimizations that CUDA has that are nowhere near supported in other software or even hardware. Custom cuda kernels also aren't as rare as one might think, they will often just be hidden unless you're looking at libraries. Our more well known example is going to be StyleGAN[0] but it isn't uncommon to see elsewhere, even in research code. Swin even has a cuda kernel[1]. Or find torch here[1] (which github reports that 4% of the code is cuda (and 42% C++ and 2% C)). These things are everywhere. I don't think pytorch and tensorflow could ever be agnostic, there will always be a difference just because you have to spend resources differently (developing kernels is time resource). We can draw evidence by looking at Intel MKL, which is still better than open source libraries and has been so for a long time.
I really do want AMD to compete in this space. I'd even love a third player like Intel. We really do need competition here, but it would be naive to think that there's going to be a quick catchup here. AMD has a lot of work to do and posting a few bounties and starting a company (idk, called "micro grad"?) isn't going to solve the problem anytime soon.
And fwiw, I'm willing to bet that most AI companies would rather run in house servers than from cloud service providers. The truth is that right now just publishing is extremely correlated to compute infrastructure (doesn't need to be but with all the noise we've just said "fuck the poor" because rejecting is easy) and anyone building products has costly infrastructure.
There are a lot of optimizations that CUDA has that are nowhere near supported in other software or even hardware. Custom cuda kernels also aren't as rare as one might think, they will often just be hidden unless you're looking at libraries. Our more well known example is going to be StyleGAN[0] but it isn't uncommon to see elsewhere, even in research code. Swin even has a cuda kernel[1]. Or find torch here[1] (which github reports that 4% of the code is cuda (and 42% C++ and 2% C)). These things are everywhere. I don't think pytorch and tensorflow could ever be agnostic, there will always be a difference just because you have to spend resources differently (developing kernels is time resource). We can draw evidence by looking at Intel MKL, which is still better than open source libraries and has been so for a long time.
I really do want AMD to compete in this space. I'd even love a third player like Intel. We really do need competition here, but it would be naive to think that there's going to be a quick catchup here. AMD has a lot of work to do and posting a few bounties and starting a company (idk, called "micro grad"?) isn't going to solve the problem anytime soon.
And fwiw, I'm willing to bet that most AI companies would rather run in house servers than from cloud service providers. The truth is that right now just publishing is extremely correlated to compute infrastructure (doesn't need to be but with all the noise we've just said "fuck the poor" because rejecting is easy) and anyone building products has costly infrastructure.
[0] https://github.com/NVlabs/stylegan2-ada-pytorch/blob/d72cc7d...
[1] https://github.com/microsoft/Swin-Transformer/blob/2cb103f2d...
[2] https://github.com/pytorch/pytorch/tree/main/aten/src