Skip to content
Documentation: GPUs and drivers

Documentation · Usage

GPUs and drivers

What the first setup installs for Intel, AMD and NVIDIA, what has been tested and how it works with more than one GPU.

Alpha

This documentation covers vGPU OS 0.3.0, in alpha. Screens and commands may change between versions; to see the installed version, use vgpu-os version.

vGPU OS does not include any video driver inside the ISO. During the first setup, it identifies each GPU in the computer and installs only what it needs — from the official Debian 13 repository and, for the newer NVIDIA cards, from NVIDIA's own repository — after showing the plan and asking for confirmation.

What gets installed

On every computer: the Vulkan loader, the Mesa Vulkan drivers, Mesa OpenGL, and the verification tools (vulkan-tools). Mesa is version 26, from Debian 13's official backports: Debian 13's Mesa 25 does not give Reims vGPU what it needs on AMD GPUs. Every image also carries nvtop, to watch any GPU's use from a terminal.

GPU Driver Vulkan Firmware
Intel i915 or xe, chosen by the kernel Mesa ANV firmware-intel-graphics and firmware-misc-nonfree
AMD amdgpu Mesa RADV firmware-amd-graphics
NVIDIA, from the GTX 16xx on NVIDIA's own driver, with its open kernel modules (nvidia-kernel-open-dkms), from NVIDIA's repository for Debian 13 nvidia-driver-libs —
Older NVIDIA Driver from Debian non-free, chosen by nvidia-detect nvidia-vulkan-icd firmware-nvidia-graphics

The NVIDIA driver is only installed if you authorize it. It comes with the driver's libcuda1 library, which NVIDIA's Vulkan opens for the ray tracing features Reims vGPU asks for; the CUDA Toolkit is never installed.

Important

From the GTX 16xx on (RTX 20, 30, 40 and 50 included), the driver comes from NVIDIA's repository, because Debian 13's driver (550 series) does not cover the newest cards and does not build for the ISO's kernel. The module is built for the computer, which takes a few minutes, and is not signed: Secure Boot must be off. If nvidia-detect finds no Debian driver for an older card, the setup offers NVIDIA's in the same way.

What has been tested

On real hardware:

  • Intel Core i9-14900K with UHD 770 integrated graphics (Raptor Lake) as the only GPU: first setup with one automatic restart, and a VM with macOS Sequoia installed and starting on its own.
  • The same computer with an AMD Radeon RX 6600 added, with the monitor sometimes on one GPU and sometimes on the other, and VMs with Sonoma, Sequoia and Tahoe running on either of them.
  • The same computer with an NVIDIA GeForce RTX 5060 added, with the driver from NVIDIA's repository.
  • Intel Core i9-12900K and AMD Ryzen 9 5950X, each with a GeForce RTX 5060, including a VM backed up on one computer and restored on the other.
  • AMD Ryzen 9 5950X with an AMD Radeon RX 6900 XT.
  • A notebook with an Intel Core i7-10750H, UHD graphics and a GeForce RTX 2060 Mobile, with NVIDIA's open driver and a VM with macOS Sequoia installed on the RTX 2060.
  • AMD Radeon RX 580, in the initial test rounds.

Compatible with Vulkan GPUs on Linux

Because Reims vGPU works on top of Vulkan, other NVIDIA, Intel Arc and AMD integrated GPUs are compatible with the project too: all of them have drivers with Vulkan support on Linux.

Reims vGPU needs the GPU's Vulkan driver to have the VK_EXT_external_memory_host extension. nouveau (the community's open driver for NVIDIA) does not, which is why an NVIDIA card without NVIDIA's own driver cannot run VMs.

Not every card and computer combination has gone through vGPU OS validation, so results may vary with the GPU model and the driver.

Graphics validation

vGPU OS does not consider a GPU ready just because the driver loaded: it checks whether it actually responds to Vulkan. The result appears at the end of the first setup and in vgpu-os hw validate:

  • READY — the GPU responded to Vulkan;
  • SOFTWARE — only software rendering is available; macOS is far too slow;
  • FAIL — the GPU did not pass; OSX-REIMS is not prepared.

The validation also says, for each GPU, whether it can run a VM (the VM memory line) and how many of the computer's GPUs can (the VM GPU line). With NVIDIA, it checks that the libcuda.so.1 library is installed (the CUDA driver line). These lines only warn: the session opens anyway, and OSX-REIMS does not offer the VMs a GPU that cannot run them.

If a GPU has its driver loaded but its firmware was not read, the setup restarts the computer once, on its own, to fix it.

More than one GPU

  • Each GPU is validated separately. The GPU the monitor is connected to must respond to Vulkan.
  • In OSX-REIMS, each VM has the GPU setting: which card it runs on. On automatic, the Reims vGPU rule applies — discrete before integrated.
  • The image is always output through the monitor's GPU, as on a hybrid laptop.
  • If the GPU chosen for a VM is no longer in the computer, the VM starts on the automatic one, with a warning.
  • vGPU OS updates what it knows about the hardware on every boot: if you move the monitor to another card, the change takes effect from the next boot, not while the computer is on.

Black screen or corrupted image

If the USB drive shows no image, use the Install vGPU OS (safe graphics) option, with basic video. For an already installed system, see Troubleshooting.