Running ComfyUI on AMD Ryzen AI Max 395 (Strix Halo)

The Challenge

AMD’s Ryzen AI Max 395 (Strix Halo) represents a new era in APU computing with its powerful integrated RDNA 3.5 GPU. However, running AI workloads like ComfyUI on this architecture requires specific ROCm configurations that aren’t yet mainstream.

The main challenge? The gfx1151 architecture isn’t fully supported in standard PyTorch ROCm builds—until now.

Our Solution

We’ve created strix-halo-comfyui, an installation toolkit that gets ComfyUI running on Strix Halo APUs with minimal friction.

Key Features

How It Works

The installation leverages a specially compiled PyTorch wheel that includes gfx1151 support:

1
2
3

git clone https://github.com/smarttechlabs-projects/strix-halo-comfyui.git
cd strix-halo-comfyui
./install.sh

The script handles:

1. Environment Setup: Creates a Python 3.11 virtual environment
2. PyTorch Installation: Installs the gfx1151-compatible PyTorch build
3. ComfyUI Clone: Downloads and configures ComfyUI
4. Dependencies: Installs all required Python packages

Requirements

Why This Matters

Strix Halo APUs offer an interesting proposition for AI workloads:

• Unified Memory: Up to 128GB accessible by both CPU and GPU
• Power Efficiency: Better performance per watt than discrete GPUs for certain workloads
• Cost Effective: No need for expensive discrete GPU for image generation

With our installation scripts, you can now tap into this potential for ComfyUI workflows without wrestling with compatibility issues.

Get Started

Check out the repository: smarttechlabs-projects/strix-halo-comfyui

We welcome contributions and feedback. If you encounter issues or have improvements, please open an issue or pull request.

Building AI solutions on AMD hardware? Contact us to discuss your project requirements.