Chapter 34. PRIME Render Offload

PRIME render offload is the ability to have an X screen rendered by one GPU, but choose certain applications within that X screen to be rendered on a different GPU. This is particularly useful in combination with dynamic power management to leave an NVIDIA GPU powered off except when it is needed to render select performance-sensitive applications.

The GPU rendering the majority of the X screen is known as the "sink", and the GPU to which certain application rendering is "offloaded" is known as the "source". The render offload source produces content that is presented on the render offload sink. The NVIDIA driver can function as a PRIME render offload source, to offload rendering of GLX+OpenGL or Vulkan, presenting to an X screen driven by the xf86-video-modesetting X driver.

X Server Requirements

NVIDIA's PRIME render offload support requires X.Org xserver version 1.20.7 or newer.

Configure the X Server

On systems with both an integrated GPU and an NVIDIA discrete GPU, the X.Org X server version 1.20.7 and newer will automatically use NVIDIA's PRIME render offload support if the system BIOS is configured to boot on the iGPU and no other explicit configuration files are present. Note that some Linux distributions (such as Ubuntu) may configure the X server differently. Please refer to your distribution's documentation for details.

If GPU screen creation was successful, the log file /var/log/Xorg.0.log should contain lines with "NVIDIA(G0)", and querying the RandR providers with xrandr --listproviders should display a provider named "NVIDIA-G0" (for "NVIDIA GPU screen 0"). For example:

Providers: number : 2
Provider 0: id: 0x221 cap: 0x9, Source Output, Sink Offload crtcs: 3 outputs: 6 associated providers: 0 name:modesetting
Provider 1: id: 0x1f8 cap: 0x0 crtcs: 0 outputs: 0 associated providers: 0 name:NVIDIA-G0

Configure Graphics Applications to Render Using the GPU Screen

To configure a graphics application to be offloaded to the NVIDIA GPU screen, set the environment variable __NV_PRIME_RENDER_OFFLOAD to 1. If the graphics application uses Vulkan or EGL, that should be all that is needed. If the graphics application uses GLX, then also set the environment variable __GLX_VENDOR_LIBRARY_NAME to nvidia, so that GLVND loads the NVIDIA GLX driver.

Examples:

__NV_PRIME_RENDER_OFFLOAD=1 vkcube
__NV_PRIME_RENDER_OFFLOAD=1 __GLX_VENDOR_LIBRARY_NAME=nvidia glxinfo | grep vendor

Finer-Grained Control of Vulkan

The __NV_PRIME_RENDER_OFFLOAD environment variable causes the special Vulkan layer VK_LAYER_NV_optimus to be loaded. Vulkan applications use the Vulkan API to enumerate the GPUs in the system and select which GPU to use; most Vulkan applications will use the first GPU reported by Vulkan. The VK_LAYER_NV_optimus layer causes the GPUs to be sorted such that the NVIDIA GPUs are enumerated first. For finer-grained control, the VK_LAYER_NV_optimus layer looks at the __VK_LAYER_NV_optimus environment variable. The value NVIDIA_only causes VK_LAYER_NV_optimus to only report NVIDIA GPUs to the Vulkan application. The value non_NVIDIA_only causes VK_LAYER_NV_optimus to only report non-NVIDIA GPUs to the Vulkan application.

Examples:

__NV_PRIME_RENDER_OFFLOAD=1 __VK_LAYER_NV_optimus=NVIDIA_only vkcube
__NV_PRIME_RENDER_OFFLOAD=1 __VK_LAYER_NV_optimus=non_NVIDIA_only vkcube

Finer-Grained Control of OpenGL

For OpenGL with either GLX or EGL, the environment variable __NV_PRIME_RENDER_OFFLOAD_PROVIDER provides finer-grained control. While __NV_PRIME_RENDER_OFFLOAD=1 tells GLX or EGL to use the first NVIDIA GPU screen, __NV_PRIME_RENDER_OFFLOAD_PROVIDER can use an RandR provider name to pick a specific NVIDIA GPU screen, using the NVIDIA GPU screen names reported by `xrandr --listproviders`.

Examples:

__NV_PRIME_RENDER_OFFLOAD=1 __GLX_VENDOR_LIBRARY_NAME=nvidia glxgears
__NV_PRIME_RENDER_OFFLOAD_PROVIDER=NVIDIA-G0 __GLX_VENDOR_LIBRARY_NAME=nvidia glxgears
__NV_PRIME_RENDER_OFFLOAD=1 eglinfo
__NV_PRIME_RENDER_OFFLOAD_PROVIDER=NVIDIA-G0 eglinfo

Troubleshooting

After starting the X server, verify that the xf86-video-modesetting X driver is using "glamoregl". The log file /var/log/Xorg.0.log should contain something like this:

[1272173.618] (II) Loading sub module "glamoregl"
[1272173.618] (II) LoadModule: "glamoregl"
[1272173.618] (II) Loading /usr/lib/xorg/modules/libglamoregl.so
[1272173.622] (II) Module glamoregl: vendor="X.Org Foundation"
[1272173.622]   compiled for 1.20.4, module version = 1.0.1
[1272173.622]   ABI class: X.Org ANSI C Emulation, version 0.4
[1272173.638] (II) modeset(0): glamor X acceleration enabled on Mesa DRI Intel(R) HD Graphics 630 (Kaby Lake GT2)
[1272173.638] (II) modeset(0): glamor initialized

If glamoregl could not be loaded, the X log may report something like:

[1271802.673] (II) Loading sub module "glamoregl"
[1271802.673] (II) LoadModule: "glamoregl"
[1271802.673] (WW) Warning, couldn't open module glamoregl
[1271802.673] (EE) modeset: Failed to load module "glamoregl" (module does not exist, 0)
[1271802.673] (EE) modeset(0): Failed to load glamor module.

in which case, consult your distribution's documentation for how to (re-)install the package containing glamoregl.

If the server didn't create a GPU screen automatically, ensure that the nvidia_drm kernel module is loaded. This should normally happen by default, but you can confirm by running lsmod | grep nvidia_drm to see if the kernel module is loaded. Run modprobe nvidia_drm to load it.

If automatic configuration does not work, it may be necessary to explicitly configure the iGPU and dGPU devices in xorg.conf:

    Section "ServerLayout"
      Identifier "layout"
      Screen 0 "iGPU"
    EndSection

    Section "Device"
      Identifier "iGPU"
      Driver "modesetting"
    EndSection

    Section "Screen"
      Identifier "iGPU"
      Device "iGPU"
    EndSection

    Section "Device"
      Identifier "dGPU"
      Driver "nvidia"
    EndSection