GPU Setup¶

Bazzite AI supports all modern GPUs through open-source drivers. This guide covers compatibility, setup, and troubleshooting.

Supported GPUs¶

Graphics Stack¶

Quick Setup¶

Configure GPU container support using ujust commands:

NVIDIA Setup¶

Supported Cards¶

• RTX 40-series (Ada Lovelace)
• RTX 30-series (Ampere)
• RTX 20-series (Turing)

Older cards (GTX 10-series and earlier) are not supported due to open driver requirements.

Verify GPU Detection¶

# On host
nvidia-smi

# Expected output shows your GPU model, driver version, CUDA version

Container GPU Access¶

This configures:

• nvidia-container-toolkit for container GPU passthrough
• CDI (Container Device Interface) config at /etc/cdi/nvidia.yaml

Verify Container GPU Access¶

Inside the container:

nvidia-smi
python -c "import torch; print(torch.cuda.is_available())"

AMD Setup¶

Supported Cards¶

• RX 7000-series (RDNA 3)
• RX 6000-series (RDNA 2)
• RX 5000-series (RDNA)
• Vega (GCN 5)
• RX 500/400-series (GCN 4)

No Setup Required¶

AMD GPUs work automatically via the AMDGPU driver and Mesa RADV.

Verify GPU Detection¶

# On host
ls /dev/dri/
# Should show: card0, renderD128

# Check GPU info
glxinfo | grep "OpenGL renderer"

Container GPU Access¶

AMD GPUs work automatically via /dev/dri passthrough. Use ujust apptainer shell to access.

Intel Setup¶

Supported GPUs¶

• Arc B-series (Battlemage)
• Arc A-series (Alchemist)
• Iris Xe (Gen 12)
• UHD Graphics (Gen 7+)

No Setup Required¶

Intel GPUs work automatically via i915/xe drivers and Mesa ANV.

Verify GPU Detection¶

# Check GPU
ls /dev/dri/
glxinfo | grep "OpenGL renderer"

# For Arc GPUs
vainfo

Container GPU Access¶

Same as AMD - automatic via /dev/dri. Use ujust apptainer shell to access.

Multi-GPU Systems¶

NVIDIA + Intel (Common Laptop Config)¶

Bazzite AI handles hybrid graphics automatically. For GPU containers:

# Force NVIDIA GPU when running Apptainer
__NV_PRIME_RENDER_OFFLOAD=1 ujust apptainer shell

Multiple NVIDIA GPUs¶

# Use specific GPU
CUDA_VISIBLE_DEVICES=0 python train.py
CUDA_VISIBLE_DEVICES=1 python train.py

# Use all GPUs
CUDA_VISIBLE_DEVICES=0,1 python train.py

Troubleshooting¶

NVIDIA: "CUDA not available"¶

1. Check driver loaded:

   lsmod | grep nvidia
   # Should show nvidia, nvidia_modeset, nvidia_uvm
   

2. Verify nvidia-smi:

   nvidia-smi
   # If fails, driver not loaded properly
   

3. Configure GPU containers:

   Command	Description
   ujust config gpu setup	Setup container toolkit
   ujust config gpu status	Verify configuration

4. Reboot and retry:

   systemctl reboot
   

NVIDIA: "Driver/library version mismatch"¶

After kernel updates, NVIDIA modules may need rebuilding:

# Check current kernel vs driver kernel
uname -r
modinfo nvidia | grep vermagic

# If mismatched, reboot to load new drivers
systemctl reboot

AMD/Intel: "No GPU found in container"¶

Ensure /dev/dri is accessible:

# Check host
ls -la /dev/dri/

Permission Denied on /dev/dri¶

Add user to video and render groups:

# On host
sudo usermod -aG video,render $USER
# Log out and back in

GPU Not Detected After Update¶

# Check OSTree status
rpm-ostree status

# If issues, rollback
rpm-ostree rollback
systemctl reboot

Low Performance¶

1. Check power mode:

   # NVIDIA
   nvidia-smi -q -d POWER
   
   # Set performance mode
   nvidia-smi -pm 1
   

2. Check thermal throttling:

   nvidia-smi dmon -s pucvmet
   

3. Verify PCIe bandwidth:

   nvidia-smi -q -d PCIE
   

Testing GPU Access¶

Quick Tests¶

# NVIDIA - on host
nvidia-smi

# All GPUs - OpenGL
glxgears

# Vulkan
vkcube

In Containers¶

# PyTorch
import torch
print(f"CUDA: {torch.cuda.is_available()}")
print(f"Devices: {torch.cuda.device_count()}")
print(f"Name: {torch.cuda.get_device_name(0)}")

# Quick benchmark
import time
x = torch.randn(10000, 10000, device='cuda')
start = time.time()
y = torch.matmul(x, x)
torch.cuda.synchronize()
print(f"Matrix multiply: {time.time() - start:.3f}s")

See Also¶

• Requirements - Hardware requirements
• nvidia-python - ML/AI development
• Apptainer Command Reference - HPC container commands
• Config Command Reference - GPU configuration commands