GPU Compat Tracker: untangling the CUDA stack

Field notes

The tracker separates compatibility into CUDA Toolkit, NVIDIA drivers, PyTorch, TensorFlow, ROCm, and GPU hardware, with a checker that starts from a framework version. Its ROCm coverage includes gfx targets, PyTorch-ROCm builds, and kernel driver minimums, while the GPU section lists VRAM, bandwidth, and FP16/FP8 compute. The site says ambiguous vendor documentation is flagged instead of resolved by guesswork, but also notes that community-maintained data can lag official releases.