[cwegener]

Yes. CUDA is still entrenched due to its ease of use when compared to specialised AI hardware SDKs.

OpenMP is still a thing in 2024, but I presume that is not the kind of scale you are asking about.

If you have MI accelerators, you'll be using ROCm anyway (although AMD contracted Andrzej Janik in 2022 to make ZLUDA run on AMD GPUs. I have no idea what the practical applications of it are at the moment.)

The only other serious challenger (apart from the existing GPU manufactures like AMD) who is trying to give Nvidia a run for its money is Tenstorrent and their TT-Buda software kit.