Skip to main content
ComfyUI supports a wide range of hardware platforms, from high-end NVIDIA GPUs to CPU-only systems. This guide covers installation and configuration for different hardware types.

​
Supported Hardware

ComfyUI works on:
  • NVIDIA GPUs (CUDA) - Best performance and compatibility
  • AMD GPUs (ROCm) - Good performance on supported cards
  • Intel GPUs (Arc, Xe) - Experimental support via oneAPI/XPU
  • Apple Silicon (M1/M2/M3/M4) - MPS acceleration support
  • Ascend NPUs - Huawei Ascend accelerators
  • Cambricon MLUs - Cambricon machine learning units
  • Iluvatar Corex - Iluvatar AI accelerators
  • CPU - Works on any system (slow)

​
NVIDIA GPUs

NVIDIA GPUs offer the best performance and compatibility with ComfyUI. Requires CUDA-capable hardware.

​
Requirements

  • NVIDIA GPU with CUDA compute capability 3.5 or higher
  • Updated NVIDIA drivers
  • Python 3.10+ (3.13 recommended)
  • PyTorch 2.4 or newer

​
Installation

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

​
Install PyTorch with CUDA 12.6 (For Older GPUs)

For NVIDIA 10-series and older GPUs: 
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu126

​
Install PyTorch Nightly (Latest Features)

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

​
Portable Windows Builds

Pre-built portable packages are available for Windows:
Update your NVIDIA drivers if ComfyUI doesn’t start. The portable builds come with Python and PyTorch pre-configured.
Best performance with all models loaded in VRAM:
python main.py --highvram --preview-method taesd
Or for maximum performance:
python main.py --gpu-only --preview-method taesd

​
Troubleshooting

​
”Torch not compiled with CUDA enabled”

Uninstall and reinstall PyTorch: 
pip uninstall torch
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130

​
CUDA malloc errors

Disable cudaMallocAsync: 
python main.py --disable-cuda-malloc

​
Out of memory errors

Reduce VRAM usage: 
python main.py --lowvram --fp16-vae

​
AMD GPUs

AMD GPUs are supported through ROCm on Linux and experimental Windows builds for RDNA 3/4. Supported on most modern AMD GPUs with ROCm.

​
Install PyTorch with ROCm 7.1 (Stable)

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.1

​
Install PyTorch with ROCm 7.2 (Nightly)

May have performance improvements: 
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.2

​
Running ComfyUI

python main.py --use-pytorch-cross-attention --preview-method auto

​
Windows (Experimental)

Experimental support for RDNA 3, 3.5, and 4 GPUs only.
Windows AMD support is experimental and has less hardware compatibility than Linux.

​
RDNA 3 (RX 7000 Series)

pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/

​
RDNA 3.5 (Strix Halo / Ryzen AI Max+ 365)

pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx1151/

​
RDNA 4 (RX 9000 Series)

pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx120X-all/

​
Portable Windows Build

Experimental AMD Portable Build

​
Unsupported AMD Cards (Linux)

For older RDNA2 or unsupported cards, use the HSA_OVERRIDE_GFX_VERSION environment variable: 
# For RX 6700, 6600, and RDNA2 cards
HSA_OVERRIDE_GFX_VERSION=10.3.0 python main.py

# For RX 7600 and RDNA3 cards
HSA_OVERRIDE_GFX_VERSION=11.0.0 python main.py

​
AMD ROCm Optimizations

Enable experimental memory-efficient attention: 
TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 python main.py --use-pytorch-cross-attention
Enable TunableOp for potential speedup (slow first run): 
PYTORCH_TUNABLEOP_ENABLED=1 python main.py --use-pytorch-cross-attention
These optimizations are already enabled by default on RDNA3 GPUs.

​
Intel GPUs

Intel Arc and Xe GPUs are supported via Intel Extension for PyTorch.

​
Requirements

  • Intel Arc GPU (A-series) or Intel Xe integrated graphics
  • Windows or Linux
  • Python 3.10+

​
Installation

​
Install PyTorch XPU (Stable)

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

​
Install PyTorch XPU (Nightly)

May have performance improvements: 
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu
See the PyTorch XPU documentation for more details.

​
Running ComfyUI

python main.py --use-pytorch-cross-attention --preview-method auto

​
Device Selection

Select specific Intel device: 
python main.py --oneapi-device-selector "level_zero:0"

​
Disable IPEX Optimizations

If you encounter issues: 
python main.py --disable-ipex-optimize

​
Apple Silicon (M1/M2/M3/M4)

ComfyUI supports Apple Silicon Macs with MPS (Metal Performance Shaders) acceleration.

​
Requirements

  • Mac with Apple Silicon (M1, M2, M3, or M4)
  • macOS 12.0 or later
  • Python 3.10+ (3.13 recommended)

​
Installation

  1. Install PyTorch with MPS support:
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cpu
See Apple’s PyTorch guide for installation instructions.
  1. Install ComfyUI dependencies:
pip install -r requirements.txt
  1. Add models to the appropriate folders as described in the installation guide.

​
Running ComfyUI

python main.py --preview-method auto
ComfyUI automatically detects and uses MPS acceleration on Apple Silicon.

​
Performance Tips

  • Unified Memory: Apple Silicon benefits from unified memory architecture
  • Model Size: Can handle larger models than GPU VRAM alone would suggest
  • Performance: Expect 2-4x slower than equivalent NVIDIA GPU, but very capable
Apple Silicon works well with:
  • SD 1.5 models
  • SDXL (on M2 Pro/Max/Ultra or M3/M4)
  • Flux (on M3 Max/Ultra or M4 Max/Ultra with 64GB+ RAM)

​
Ascend NPUs

Huawei Ascend NPUs are supported via Ascend Extension for PyTorch.

​
Requirements

  • Ascend NPU hardware
  • Supported Linux kernel (see Ascend documentation)
  • Ascend Basekit (driver, firmware, CANN)

​
Installation

  1. Install the recommended Linux kernel version from the torch-npu installation page
  2. Install Ascend Basekit (driver, firmware, and CANN) following platform-specific instructions
  3. Install torch-npu packages following the installation guide
  4. Install ComfyUI dependencies: 
pip install -r requirements.txt

​
Running ComfyUI

python main.py
ComfyUI automatically detects and uses Ascend NPUs.

​
Cambricon MLUs

Cambricon Machine Learning Units are supported via Cambricon Extension for PyTorch.

​
Requirements

  • Cambricon MLU hardware
  • Cambricon CNToolkit
  • PyTorch with torch_mlu

​
Installation

  1. Install Cambricon CNToolkit following the installation guide
  2. Install PyTorch with torch_mlu support following the user guide
  3. Install ComfyUI dependencies: 
pip install -r requirements.txt

​
Running ComfyUI

python main.py

​
Iluvatar Corex

Iluvatar AI accelerators are supported via Iluvatar Extension for PyTorch.

​
Requirements

  • Iluvatar Corex hardware
  • Iluvatar Corex Toolkit

​
Installation

  1. Install Iluvatar Corex Toolkit following the installation guide
  2. Install ComfyUI dependencies: 
pip install -r requirements.txt

​
Running ComfyUI

python main.py

​
CPU Only

ComfyUI can run on CPU alone, though it will be significantly slower than GPU acceleration.

​
Requirements

  • Any x86_64 or ARM CPU
  • 16GB+ RAM recommended (32GB+ for larger models)
  • Python 3.10+

​
Installation

Install PyTorch CPU version: 
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

​
Running ComfyUI

python main.py --cpu --preview-method latent2rgb

​
Performance Expectations

  • SD 1.5: 2-5 minutes per image on modern CPUs
  • SDXL: 10-30 minutes per image
  • Flux: Not recommended (extremely slow)

​
Optimization Tips

python main.py --cpu --lowvram --fp16-unet --preview-method latent2rgb

​
Hardware Selection Guide

​
Which GPU Should I Buy?

See the ComfyUI GPU recommendations wiki for detailed buying advice.

​
Quick Recommendations

Budget

NVIDIA RTX 3060 12GB or RTX 4060 Ti 16GBGood for SD 1.5, SDXL, and smaller models.

Enthusiast

NVIDIA RTX 4070 Ti Super 16GB or RTX 4080 16GBGreat for SDXL, Flux schnell, and most workflows.

Professional

NVIDIA RTX 4090 24GB or RTX 6000 Ada 48GBBest for all models including Flux dev, video generation, and heavy workflows.

Server/Multi-GPU

NVIDIA A6000 48GB or H100 80GBEnterprise solutions for maximum performance and multi-user setups.

​
VRAM Requirements by Model

Model TypeMinimum VRAMRecommended VRAM
SD 1.54GB8GB
SDXL6GB12GB
Flux Schnell12GB16GB
Flux Dev16GB24GB
Video (SVD)12GB24GB
Mochi20GB32GB
Hunyuan Video32GB60GB
Lower VRAM amounts can work with --lowvram, --novram, and FP8 quantization, but will be slower.

​
Multi-GPU Setup

ComfyUI supports multi-GPU systems.

​
Select Primary GPU

# Use GPU 0 as primary, hide other GPUs
python main.py --cuda-device 0

# Use GPU 1 as primary, keep others visible
python main.py --default-device 1

​
Environment Variables

Alternatively, set CUDA device visibility: 
export CUDA_VISIBLE_DEVICES=0,1
python main.py

​
Benchmarking Your Hardware

To test your hardware performance:
  1. Load a standard workflow (e.g., SDXL text-to-image)
  2. Run with different settings and note generation times:
# Test 1: Default
python main.py

# Test 2: High VRAM
python main.py --highvram

# Test 3: Optimizations
python main.py --use-pytorch-cross-attention

# Test 4: Experimental
python main.py --fast
  1. Compare execution times in the console output

​
Getting Help

If you encounter hardware-specific issues: Provide details about:
  • Your hardware (GPU model, VRAM, CPU, RAM)
  • Operating system and version
  • PyTorch version (python -c "import torch; print(torch.__version__)")
  • Full error messages and console output

Build docs developers (and LLMs) love

Get started for freeTalk to us