Home

If you find any mistakes here or anywhere else on my website, email website@CoryHufford.com.

Selecting Which GPU to Use for OpenGL and Vulkan Programs in Linux

September 30, 2026

On a Linux system with multiple GPUs, which GPU is used for a particular OpenGL or Vulkan process can be controlled with the DRI_PRIME environment variable (with some exceptions and caveats), which is documented here. Since this is a Mesa feature, this works with all discrete and integrated AMD, Intel, and NVIDIA GPUs as long as the open source Mesa drivers are being used.

The DRI_PRIME environment variable can take several types of GPU identifiers; I find the most convenient to be the PCI vendor and device ID, since that's the only type that's stable across machines and reboots. Note that PCI device IDs can be ambiguous if there are multiple GPUs of the same model or two closely related models in the same system; in that case, a different form of identifier would have to be used for DRI_PRIME. The GPUs present on a system can be listed with their PCI IDs with lspci -nn | grep VGA. For example, my current desktop machine with an AMD Ryzen 7950X (which has an integrated GPU) and an AMD RX 7900 XTX discrete GPU lists:

03:00.0 VGA compatible controller [0300]: Advanced Micro Devices, Inc. [AMD/ATI] Navi 31 [Radeon RX 7900 XT/7900 XTX/7900 GRE/7900M] [1002:744c] (rev c8)
18:00.0 VGA compatible controller [0300]: Advanced Micro Devices, Inc. [AMD/ATI] Raphael [1002:164e] (rev c1)

[1002:744c] is the vendor and device ID of the 7900 XTX and [1002:164e] is the vendor and device ID for the 7950X's integrated GPU. The PCI ID Repository is a useful resource for finding or identifying PCI IDs. A selection of discrete and integrated GPUs and their PCI IDs are shown in Table 1:

Using vendor ID 1a2b and device ID 3c4d as an example, the GPU can be specified with DRI_PRIME=1a2b:3c4d!. Note the trailing exclamation mark - it's often required for Vulkan programs. The OpenGL driver chooses the GPU on behalf of the program and always respects DRI_PRIME with or without an exclamation mark. The Vulkan driver, however, presents programs with a list of GPUs to choose from. Some Vulkan programs choose the first GPU in the list, others choose based on their own criteria. Without the exclamation mark, DRI_PRIME causes the specified GPU to be moved to the beginning of the list, but the program is still free to choose any GPU from the list. With the exclamation mark, all GPUs are removed from the list except the specified one, forcing the program to use it. This GPU sorting/filtering is implemented in a Vulkan layer outside of the driver itself, so the layer has to be installed for DRI_PRIME to have any effect on Vulkan programs. The layer is called libVkLayer_MESA_device_select.so, and is often included with the normal Mesa Vulkan drivers package (usually called mesa-vulkan-drivers); Arch Linux is a notable exception, where it's provided by the vulkan-mesa-implicit-layers package. The concept of Vulkan layers is briefly described in the Vulkan Loader documentation here.

To test that GPU selection is working, any graphics-intensive program can be used to create a GPU load that can be seen in monitoring software like btop. There are graphics test programs specifically for OpenGL and Vulkan that can be used for this:

OpenGL

glxgears -fullscreen opens a window that renders a set of spinning gears. -fullscreen produces a larger (more noticeable) GPU load. glxgears is provided by the mesa-utils package in Debian- and Arch-family distributions and glx-utils in Fedora-family distributions.

Vulkan

vkcube --present_mode 0 opens a window that renders a spinning cube. --present_mode 0 disables frame synchronization to let the program run at a higher frame rate, producing a more noticeable GPU load. vkcube is provided by the vulkan-tools package in most major distributions.

Back to top

Home