Which CUDA / PyTorch / ROCm versions actually work together?
A single reference for GPU software-stack compatibility — CUDA Toolkit, NVIDIA drivers, PyTorch, and ROCm — including the AMD/ROCm side most compatibility trackers skip entirely.
Driver minimums, supported architectures, cuDNN compatibility.
Supported GPU targets (gfx), PyTorch-ROCm builds, kernel driver minimums.
Official CUDA, ROCm, and Python support per release.
Official CUDA, cuDNN, and Python support per release.
VRAM, bandwidth, and FP16/FP8 compute for NVIDIA, AMD, and Intel GPUs, linked to CUDA/ROCm support.
Put up to 4 GPUs side by side across vendors.
Pick a framework version, get the CUDA/ROCm versions that work with it.
Pick an LLM, quantization, and context length — see the VRAM it needs and which GPUs can run it.
Plain-language definitions for VRAM, CUDA, ROCm, TFLOPS, quantization, and every other term used on this site.