Why is ComfyUI not using the GPU?

last edited

ComfyUI is not using the GPU when the installed PyTorch version lacks a matching GPU backend or when a `--cpu` flag starts it. A test is `python -c "import torch; print(torch.cuda.is_available())"`, run in the interpreter that starts ComfyUI. A result of false means torch cannot see the GPU, so the setup is at fault, not the hardware.

Which causes and fixes cover most GPU problems?

causesignfix
torch without GPU supporterror: not compiled with CUDA enableduninstall torch, reinstall the GPU wheel
stale versioncu126 or cu121 on a 20-series or newer GPUswitch to cu130, update the NVIDIA driver
wrong shortcutrun_cpu.bat started the appstart run_nvidia_gpu.bat instead
radeon without rocmfailure on a radeon GPUinstall the rocm7.2 wheel or AMD's package
apple siliconmps.is_available() returns falseinstall current PyTorch, rerun main.py

The ComfyUI README requires a cu130 or newer PyTorch version on every NVIDIA GPU from the 20-series up. The portable package for 10-series GPUs ships cu126 and must skip newer GPUs. Related: credit prompts.

Do RTX 50-series, AMD and Intel GPUs need other installs?

The Blackwell notes in the ComfyUI docs say RTX 50-series GPUs need a CUDA 12.8 or newer release. Builds for cu121 fail on them. AMD GPUs need ROCm 7.2, so a CUDA package never reaches that GPU.

Unsupported RDNA2 models set `HSA_OVERRIDE_GFX_VERSION=10.3.0`. RDNA3 models set 11.0.0. An Intel GPU in PyTorch needs the XPU package, not CUDA 13.0. Macs use MPS, set up through Apple's PyTorch guide. A failed MPS check leaves the GPU idle.

Why does the task manager show little GPU use while generating?

The default task-manager GPU graph tracks the 3D engine. PyTorch compute appears under the cuda engine in the dropdown. The `nvidia-smi` command reports real GPU load. Models larger than VRAM stream weights from system RAM, so GPU load dips between transfers. Smaller formats help models fit.

Sources