ComfyUI “No CUDA Cores Available” on AMD RDNA 3.5 GPUs

ComfyUI reports “no CUDA cores available” when launched on AMD RX 395+ (RDNA 3.5) GPUs because ComfyUI defaults to NVIDIA’s CUDA backend. AMD GPUs require either the ROCm (ROCm 5.6+ for RDNA 3.5) or DirectML backend. The error occurs when neither backend is properly configured.

Fix 1: Install ROCm-Compatible PyTorch

RDNA 3.5 GPUs require ROCm 5.6 or later with the appropriate PyTorch build.

# Uninstall existing PyTorch
pip uninstall torch torchvision -y

# Install ROCm-compatible PyTorch (verify ROCm 5.6+ installed)
pip install torch torchvision --index-url https://download.pytorch.org/whl/rocm5.6

Verify ROCm detects your GPU:

rocminfo | grep -A 10 "gfx1100"  # RDNA 3.5 uses gfx1100 architecture

Fix 2: Set ComfyUI Launch Environment Variables

Configure ComfyUI to use the AMD backend via environment variables:

# Set device to use HIP (AMD's CUDA equivalent)
export PYTORCH_CUDA_ALLOCATOR=hip
export HIP_VISIBLE_DEVICES=0

# Launch ComfyUI
python main.py --amd

Alternatively, add to your ComfyUI web/index.html or create a launch script:

# Create launch_amd.sh
#!/bin/bash
export HSA_OVERRIDE_GFX_VERSION=11.0.0
export PYTORCH_ENABLE_MPS_FALLBACK=1
python main.py --amd --force-fp16

Fix 3: Use DirectML Backend (Windows)

On Windows, DirectML provides hardware acceleration without ROCm:

# Install DirectML-enabled PyTorch
pip uninstall torch torchvision -y
pip install torch torchvision directml --index-url https://download.pytorch.org/whl/directml

# Launch with DirectML flag
python main.py --directml

Note: DirectML is slower than ROC