It was my understanding that a big part of what makes CUDA so good is all the math libraries that Nvidia wrote and optimized for their GPUs. Are those libraries licensed to run on non-Nvidia GPUs?
To me, it seems like what AMD needs to do is invest heavily in the ROCm ecosystem to bring it up to the level of CUDA. But yeah, it probably would take a long time to catch up.
You have to hand it to Nvidia here. They invested in the GPGPU vision long before it was economically viable. Their long term focus has allowed them to reap incredible profits from AI now and crypto before it.
Contrast this with the Intel approach of halfheartedly trying a new market every couple years and abandoning it after itβs not immediately successful. See Larrabee for the relevant example. But also Optane and their 5G cell modem as other notables.
If you want to use both, you can write SYCL.
The worse part is that they had an internal project that could've enabled CUDA applications to run with ROCm, but they stopped funding it [1]
This is all to say - AMD GPUs make sense if you're a gamer and if they're cheaper than an NVIDIA where-ever you buy your GPUs. But otherwise, it feels like buying a SymbianOS phone while Android and iOS were into their second or third generations.
1 - https://www.xda-developers.com/nvidia-cuda-amd-zluda/#:%7E:t...