GPU Compatibility

NVIDIA GPUs

NVIDIA GPUs with CUDA are the best supported option for ComfyUI.

Supported Cards

Generation	Cards	CUDA Compute	Notes
RTX 50 series	5070 Ti, 5080, 5090	12.0	Requires CUDA 12.8+, latest PyTorch
RTX 40 series	4060, 4070, 4080, 4090	8.9	Excellent. Best price/performance
RTX 30 series	3060, 3070, 3080, 3090	8.6	Great. RTX 3060 12GB is very popular
RTX 20 series	2060, 2070, 2080	7.5	Supported but showing age
GTX 16 series	1660, 1650	7.5	Works with `--lowvram` for basic models
GTX 10 series	1080, 1070, 1060	6.1	Limited. SD 1.5 only with `--lowvram`

VRAM Recommendations

VRAM	Image Models	Video Models
4 GB	SD 1.5 (512×512, `--lowvram`)	Not practical
6 GB	SD 1.5, some SDXL	Not practical
8 GB	SDXL, Flux (GGUF Q4)	Very limited
12 GB	Flux (GGUF Q8), all image models	Short clips with quantized models
16 GB	Everything at high res	Wan 2.1, HunyuanVideo (short)
24 GB	Everything, large batches	Full video generation

Driver & CUDA & PyTorch Version Matching

Getting the right combination of driver, CUDA, and PyTorch is critical. Here's the compatibility matrix:

NVIDIA Driver	CUDA Version	PyTorch	Best For
530+	CUDA 12.1	torch 2.4.x	Older stable builds
560+	CUDA 12.6	torch 2.6.x	RTX 20/30/40 series
570+	CUDA 12.8	torch 2.8.0	RTX 50 series minimum
575+	CUDA 13.0	torch 2.9.1	RTX 50 series recommended

How to check your driver version

Run nvidia-smi in Command Prompt. The driver version is shown in the top-right corner. If the command is not found, you need to install or update your NVIDIA drivers.

Always keep your NVIDIA driver up to date: nvidia.com/Download

AMD GPUs

Windows (DirectML)

AMD GPU support on Windows uses DirectML. Use the Portable Package with the --directml flag. It works but has limitations:

Performance: Slower than NVIDIA CUDA for the same tier GPU
Compatibility: Some custom nodes only support CUDA
Setup: Use the portable package and add --directml flag

Supported cards: RX 6000 series and newer (RDNA2+)

Linux (ROCm)

AMD on Linux uses ROCm, which offers better performance than DirectML. See Manual Install for ROCm PyTorch setup:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.2.4

Supported cards: RX 6000, RX 7000 series (RDNA2/RDNA3)

Intel GPUs

Intel Arc (XPU)

Intel Arc GPUs have experimental support via the XPU backend:

Supported: Arc A770, A750, A580
Status: Experimental — expect some nodes to not work
Setup: Requires Intel oneAPI toolkit and PyTorch with XPU support

Intel Integrated Graphics

Not recommended. Shared memory is too slow and limited for practical use.

Apple Silicon (macOS)

M1, M2, M3, and M4 chips are supported via the MPS (Metal Performance Shaders) backend:

Works with unified memory (shared between CPU and GPU)
M1/M2 with 16 GB: can run SD 1.5 and SDXL
M3/M4 with 32 GB+: can run Flux and some video models
Performance is lower than a dedicated NVIDIA GPU

CPU Only

ComfyUI can run on CPU with the --cpu flag. This is very slow and only useful for testing:

SD 1.5 at 512×512: ~5-10 minutes per image
SDXL: ~15-30 minutes per image
Video generation: not practical

RTX 50 Series Setup Guide

The RTX 50 series (5070 Ti, 5080, 5090) is too new for many existing guides. Here's what you need:

Update NVIDIA drivers to the latest version (575+)
Install PyTorch with CUDA 13.0:

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130

Known issues:
- SageAttention may fail to compile — use --use-pytorch-cross-attention as a workaround
- Some older custom nodes need updates for CUDA 12.8+ compatibility
- Nunchaku plugin requires specific configuration

If you're stuck with a new 50-series GPU, our remote fix service has resolved dozens of these cases — typically in about 30 minutes. See also Common Issues — RTX 50 Series for more troubleshooting tips.

System Requirements — Full hardware and software requirements
Desktop App — Easiest install (NVIDIA only)
Portable Package — Supports NVIDIA, AMD, and CPU
Manual Install — Full control, all GPU types
Common Issues — Solutions for GPU-specific problems
Dependency Conflicts — When PyTorch CUDA versions clash