I'm trying to use one Windows PC as an all-purpose machine that stays running 24/7. Windows will handle gaming and everyday tasks, while Docker containers inside WSL2 will run Frigate, Jellyfin, Home Assistant, and Immich. The system has an ASUS ROG Strix Z890-E motherboard, an Intel Core Ultra 5 250K, a 1TB Kingston NVMe boot drive, and a 1TB WD Black SN850X for WSL2 data. For now, it has two older ASUS Strix GTX 1070 Ti graphics cards. I want the GPU in slot 1 reserved for Windows and gaming, and the GPU in slot 2 reserved for the WSL2 containers. However, WSL2 detects both GPUs and appears to combine or expose them together. Even when I specify the UUID of the second card in my Docker Compose configuration, both GPUs are still available. Is there a way to make WSL2 expose only the second GPU without using Proxmox or Unraid?
1 Answer
WSL2 doesn’t provide true PCI passthrough or a built-in way to hide one host GPU. It uses GPU paravirtualization through /dev/dxg, so WSL2 generally sees every compatible graphics adapter that Windows exposes. CUDA_VISIBLE_DEVICES can limit what CUDA-based applications use, but it doesn’t remove the other card from WSL2 entirely. Also, a GPU UUID in Compose may not affect applications such as Frigate or Jellyfin if they select hardware acceleration through VA-API, NVENC, or another device picker. You’ll likely need to select the correct device inside each container or application.
Thanks, that clears up why the UUID setting wasn’t controlling every application.
For NVIDIA containers, you can try reserving the second card by its Docker device index, for example with `device_ids: ["1"]` under the NVIDIA GPU reservation. That limits the GPU resources assigned to that container, although the application itself may still need its own encoder or acceleration setting configured.