Another factor is your system libs. An old PyTorch/CUDA combo, for instance, can hold 4000 GPUs back. I have been quite happy with Arch's build of PyTorch (whi

Another factor is your system libs. An old PyTorch/CUDA combo, for instance, can hold 4000 GPUs back. I have been quite happy with Arch's build of PyTorch (which targets CUDA 12.1 at the moment).