Graphics (GPU)

Marchyo supports Intel, AMD, and NVIDIA GPUs, including machines with several GPUs and hybrid graphics with NVIDIA PRIME.

GPU vendors

Specify which GPU vendors are present in your system:

marchyo.graphics.vendors = [ "intel" ];

A machine with several GPUs lists them all and names the one that drives the display:

marchyo.graphics = {
  vendors = [ "intel" "amd" ];
  primary = "amd";
};

Note

Find your GPUs with: lspci | grep -E 'VGA|3D|Display'

Vendor	Description
"intel"	Intel integrated graphics (iGPU)
"amd"	AMD GPUs (integrated or discrete)
"nvidia"	NVIDIA discrete GPUs

When vendors is empty (the default), no vendor-specific drivers are installed and no driver is forced; Mesa’s generic stack still runs. hardware.graphics.enable follows marchyo.desktop.enable (or any listed vendor), so a headless host without a listed GPU skips the graphics stack.

Several GPUs

Every listed vendor’s drivers, firmware and video-acceleration packages are installed. marchyo.graphics.primary picks the GPU that drives the display:

• It defaults to the only entry of vendors. On a PRIME laptop with one integrated GPU it defaults to "nvidia" in sync mode and to the integrated GPU otherwise. Any other combination of several vendors must set it; evaluation fails without it, or when it is not in vendors.
• Only the primary GPU sets the global driver selection: LIBVA_DRIVER_NAME (iHD, radeonsi, or nvidia on Turing or newer), VDPAU_DRIVER (radeonsi), and NVIDIA’s __GLX_VENDOR_LIBRARY_NAME / GBM_BACKEND.
• Per-app offload variables are not set globally. Run one app on another GPU with nvidia-offload <command> (PRIME offload mode) or DRI_PRIME=1 <command> (Mesa GPUs).
• Hyprland renders on the primary GPU first. Each listed GPU gets a stable node, /dev/dri/marchyo-<vendor> (a udev rule matching the PCI address from prime.<vendor>BusId when set, else the kernel driver), and AQ_DRM_DEVICES lists them with the primary first, because /dev/dri/card<N> numbering changes between boots.

NVIDIA options

• marchyo.graphics.nvidia.generation ("turing-or-newer" or "pre-turing", default "turing-or-newer"): the GPU generation. Turing is RTX 20xx and GTX 16xx. Pre-Turing is Maxwell, Pascal and Volta.

marchyo sets these NVIDIA options with lib.mkDefault; set them directly to override:

• hardware.nvidia.package: boot.kernelPackages.nvidiaPackages.stable, or legacy_580 for pre-Turing GPUs.
• hardware.nvidia.open: true for Turing or newer, false for pre-Turing. RTX 50xx requires the open modules.
• hardware.nvidia-container-toolkit.enable: true with marchyo.development.enable, so containers get the GPU through CDI (podman run --device nvidia.com/gpu=all ...).

Experimental power management is hardware.nvidia.powerManagement.enable (off).

marchyo.graphics.nvidia.generation = "turing-or-newer";
hardware.nvidia.package = config.boot.kernelPackages.nvidiaPackages.beta;

Caution

hardware.nvidia.open = true with generation = "pre-turing" fails evaluation. The open modules support only Turing and newer.

NVIDIA environment variables

• NVD_BACKEND is direct on Turing or newer and egl on pre-Turing.
• __GLX_VENDOR_LIBRARY_NAME=nvidia and GBM_BACKEND=nvidia-drm are set only when NVIDIA is primary: NVIDIA is the only GPU, or PRIME runs in sync mode.
• LIBVA_DRIVER_NAME=nvidia is set under the same condition, and only on Turing or newer.

In PRIME offload and reverse-sync modes the integrated GPU is primary, so these variables stay unset. Forcing them there breaks video in Chromium-based browsers.

AMD options

• marchyo.graphics.amd.vaapi.enable (bool, default true): Mesa radeonsi VA-API and VDPAU, plus vainfo. LIBVA_DRIVER_NAME=radeonsi and VDPAU_DRIVER=radeonsi are set only when AMD is primary.
• Overclocking and undervolting (LACT, CoreCtrl) need hardware.amdgpu.overdrive.enable = true;.
• marchyo.graphics.amd.rocm.enable (bool, default false): the ROCm compute stack. It installs the HIP runtime (rocmPackages.clr), links it to /opt/rocm/hip, and adds Marchyo users to the render and video groups. It requires "amd" in vendors.
• marchyo.graphics.amd.rocm.hsaOverrideGfxVersion (string or null, default null): the value of HSA_OVERRIDE_GFX_VERSION, for GPUs that ROCm does not officially support (for example "10.3.0" for RDNA2 or "11.0.0" for RDNA3).

marchyo.graphics = {
  vendors = [ "amd" ];
  amd.rocm = {
    enable = true;
    hsaOverrideGfxVersion = "11.0.0";
  };
};

NVIDIA PRIME (hybrid graphics)

For laptops with both an integrated GPU and a discrete NVIDIA GPU. With one integrated GPU, primary follows the mode:

marchyo.graphics = {
  vendors = [ "intel" "nvidia" ];

  prime = {
    enable = true;
    intelBusId = "PCI:0:2:0";
    nvidiaBusId = "PCI:1:0:0";
    mode = "offload";
  };
};

Finding bus IDs

Run lspci | grep -E 'VGA|3D' and convert hex bus IDs to the PCI:bus:device:function format (convert hex values to decimal).

PRIME options

Option	Type	Default	Description
marchyo.graphics.prime.enable	bool	false	Enable NVIDIA PRIME
marchyo.graphics.prime.intelBusId	string	""	PCI bus ID of the Intel GPU
marchyo.graphics.prime.amdgpuBusId	string	""	PCI bus ID of the AMD GPU
marchyo.graphics.prime.nvidiaBusId	string	""	PCI bus ID of the NVIDIA GPU
marchyo.graphics.prime.mode	string	"offload"	PRIME render mode

PRIME modes

Mode	Description
"offload"	On-demand rendering (default). Power efficient — use nvidia-offload <command> to run apps on the discrete GPU.
"sync"	Always use discrete GPU. Best performance but more power consumption. primary is "nvidia".
"reverse-sync"	iGPU for display, dGPU for compute.

PRIME requirements

Enabling PRIME is validated with assertions — the build fails with a clear message if any of these are not met:

• vendors must contain "nvidia" and an iGPU vendor ("intel" or "amd")
• primary must be "nvidia" in sync mode, and the integrated GPU otherwise
• prime.nvidiaBusId must be set
• the bus ID of the integrated GPU PRIME pairs with must be set: primary, or in sync mode the only integrated GPU (Intel when both are listed and intelBusId is set)

With Intel, AMD and NVIDIA all listed, PRIME pairs with the integrated GPU named by primary.

Integrated-only boot entry

On a hybrid laptop, set prime.integratedOnlySpecialisation = true to add an integrated-only boot specialisation:

marchyo.graphics.prime.integratedOnlySpecialisation = true;

The specialisation turns off PRIME, blacklists the nvidia, nvidia_modeset, nvidia_uvm, nvidia_drm and nouveau kernel modules, and removes the NVIDIA environment variables and the container toolkit. Use it to save battery when you do not need the discrete GPU.

Select the integrated-only entry in the systemd-boot menu. When the menu is hidden (boot.loader.timeout = 0), hold Space while the machine boots to show it. The option requires prime.enable.

Examples

Intel-only desktop

marchyo.graphics.vendors = [ "intel" ];

AMD-only desktop

marchyo.graphics.vendors = [ "amd" ];

NVIDIA desktop

marchyo.graphics = {
  vendors = [ "nvidia" ];
};

Intel + NVIDIA laptop (offload)

marchyo.graphics = {
  vendors = [ "intel" "nvidia" ];
  prime = {
    enable = true;
    intelBusId = "PCI:0:2:0";
    nvidiaBusId = "PCI:1:0:0";
    mode = "offload";
  };
};

Intel + AMD desktop (AMD drives the display)

marchyo.graphics = {
  vendors = [ "intel" "amd" ];
  primary = "amd";
};

Intel + AMD + NVIDIA laptop (AMD iGPU, offload)

marchyo.graphics = {
  vendors = [ "intel" "amd" "nvidia" ];
  primary = "amd";
  prime = {
    enable = true;
    amdgpuBusId = "PCI:6:0:0";
    nvidiaBusId = "PCI:1:0:0";
  };
};

AMD + NVIDIA laptop (sync)

marchyo.graphics = {
  vendors = [ "amd" "nvidia" ];
  prime = {
    enable = true;
    amdgpuBusId = "PCI:6:0:0";
    nvidiaBusId = "PCI:1:0:0";
    mode = "sync";
  };
};