Minimum driver for CUDA rows comes from the
NVIDIA CUDA release notes.
Thanks to minor-version compatibility, a driver new enough for CUDA 11.x/12.x/13.x runs every toolkit of the same major version.
ROCm GPU lists are the architectures officially supported per
AMD's documentation;
more consumer GPUs often work unofficially (e.g. via HSA_OVERRIDE_GFX_VERSION).
GPU filter compares against toolkit-level support — official wheels are sometimes built for fewer
architectures (TORCH_CUDA_ARCH_LIST), so treat results as an upper bound and check with
torch.cuda.is_available().
macOS CPU wheels include MPS (Metal) acceleration on Apple Silicon for PyTorch ≥ 1.12.
NVIDIA Jetson (Xavier, Orin) is Linux aarch64 only and uses
NVIDIA's JetPack-specific wheels —
the aarch64 CUDA wheels here target SBSA servers (e.g. GH200).