Offload rendering to Thunderbolt GPU

Offloading rendering to a Thunderbolt eGPU (NVIDIA)

Use PRIME Render Offload to render on an external NVIDIA GPU connected via Thunderbolt, while displaying on the internal screen.

Goal

Render apps on the external NVIDIA GPU and display the frames on the integrated GPU without switching the whole X session to NVIDIA.

Prerequisites

• Linux with Xorg or Wayland session that supports PRIME Render Offload
• NVIDIA proprietary driver installed and loaded
• nvidia-prime or PRIME offload support available in the driver stack
• Thunderbolt eGPU recognized by the system (lspci | grep -i nvidia)

Quick command (one-off)

Run an app on the eGPU with PRIME offload:

__NV_PRIME_RENDER_OFFLOAD=1 __GLX_VENDOR_LIBRARY_NAME=nvidia <your_command>

Examples:

__NV_PRIME_RENDER_OFFLOAD=1 __GLX_VENDOR_LIBRARY_NAME=nvidia glxinfo -B
__NV_PRIME_RENDER_OFFLOAD=1 __GLX_VENDOR_LIBRARY_NAME=nvidia firefox

On Wayland with Vulkan apps, add the Vulkan layer variable:

__NV_PRIME_RENDER_OFFLOAD=1 __GLX_VENDOR_LIBRARY_NAME=nvidia __VK_LAYER_NV_optimus=NVIDIA_only <your_command>

Verify it is using the eGPU

• OpenGL: glxinfo -B should show OpenGL renderer string: NVIDIA ...
• Vulkan: vulkaninfo | grep -A2 "GPU id" and confirm the selected device is NVIDIA
• NVIDIA state: nvidia-smi should list the process under Processes

Optional: create a helper

Add a small wrapper to run anything on the eGPU.

Create ~/bin/egpu-run and make it executable:

mkdir -p ~/bin
cat > ~/bin/egpu-run <<'EOF'
#!/usr/bin/env bash
export __NV_PRIME_RENDER_OFFLOAD=1
export __GLX_VENDOR_LIBRARY_NAME=nvidia
export __VK_LAYER_NV_optimus=NVIDIA_only
exec "$@"
EOF
chmod +x ~/bin/egpu-run

Usage:

egpu-run blender
egpu-run steam

Troubleshooting

• Driver not loaded: lsmod | grep nvidia should show modules. Reinstall or load the driver if missing.
• Library vendor mismatch: ensure __GLX_VENDOR_LIBRARY_NAME=nvidia is set for OpenGL apps.
• Wayland session: some compositors restrict offloading for XWayland or native Wayland apps. Try an Xorg session to compare.
• PRIME unavailable: check nvidia-settings under PRIME Profiles for Offloading support, or inspect /var/log/Xorg.0.log for PRIME lines.
• No processes in nvidia-smi: the app may be CPU-bound or rendering on the iGPU. Re-run with the env vars and verify via glxinfo -B.
• Thunderbolt permissions: some distros require authorizing the eGPU in the firmware or via boltctl: boltctl authorize <UUID>.

Notes

• This offloads rendering per-application. It does not switch the whole desktop to NVIDIA.
• Performance depends on the Thunderbolt bandwidth and the app's rendering path.