They don’t really need to.

Nvidia will have to validate and launch all of it’s PCIe4 on Epyc/Ryzen processors, so it’s not like AMD won’t benefit from the Deep Learning hype

Isn't HIP+ROCm a serious "competitor" to CUDA? You can convert your code automatic from CUDA to HIP. At least that's why the advertising says; I haven't used it myself. Plus, PyTorch and Tensorflow have AMD support, I thought.

It may convert your own code, but it can’t convert the CUDA libraries.

It also doesn’t help that AMD has stopped supporting ROCm for the current consumer GPUs.

Which CUDA libraries are you referring to? NVIDIA libraries like cuBLAS? There are ROCm libraries for a subset of those, but it's definitely a work-in-progress.

They might not have to, as others have pointed out. Intel, on the other hand, is fighting the multi-front battle against a lot of competitors. It’s great for consumers, but Intel might have to decide to focus too.