# Hybrid graphics on Linux: Intel iGPU plus NVIDIA eGPU for headless compute > Running an Intel integrated GPU alongside an NVIDIA Thunderbolt eGPU used only for compute on Linux — xe versus i915 binding on Arrow Lake, how the desktop picks a primary GPU, keeping GNOME off the e Parent: [Thunderbolt eGPU on Linux for local LLM inference](https://llms-explorer.com/tree/thunderbolt-egpu-linux/) · 11 facets · 63 facts · page: https://llms-explorer.com/tree/hybrid-graphics-igpu-nvidia-egpu-headless/ ## Hybrid graphics on Linux - Intel iGPU desktop plus NVIDIA Thunderbolt eGPU for headless compute - Running an Intel integrated GPU alongside an NVIDIA Thunderbolt eGPU used only for compute on Linux - xe versus i915 binding on Arrow Lake, how the desktop picks a primary GPU, keeping GNOME off the eGPU, PRIME render offload versus pure CUDA, and sharing one 16 GB GPU between several compute servers. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#hybrid-graphics-on-linux---intel-igpu-desktop-plus-nvidia-thunderbolt-egpu-for-headless-compute) - verified-as-of: 2026-09-25 Tags: [SOURCED url] = read/returned by a fetched page or search result during the research pass; [INFERRED] = reasoned from mechanism, verify on your box; [BOX] = observed on the reference machine (Intel NUC 15 Pro, Arrow Lake-P iGPU at 00:02.0, Ubuntu 26.04.1 GNOME Wayland, kernel 7.0.0-34, RTX 5080 over TB4, nvidia 610.57.04-open loaded late). — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#hybrid-graphics-on-linux---intel-igpu-desktop-plus-nvidia-thunderbolt-egpu-for-headless-compute) ## Core Concepts - Two independent GPU roles. A display/compositor GPU owns scanout and the desktop; a compute GPU only runs CUDA/Vulkan-compute. In a hybrid box the goal is usually: iGPU = display, eGPU = compute only. Arch's eGPU page states compute-only workloads (CUDA) that display nothing "should work without any extra configuration", and monitors on the iGPU work out of the box [SOURCED https://wiki.archlinux.org/title/External_GPU via search result]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#core-concepts) - Kernel driver vs userspace stack. xe/i915 (kernel DRM drivers for Intel) and nvidia/nvidia_drm (NVIDIA) are separate kernel stacks; CUDA talks to the nvidia module through /dev/nvidia*, not through DRM. DRM nodes (/dev/dri/cardN, renderDN) matter only to compositors, Vulkan/EGL/GL and PRIME. [INFERRED] — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#core-concepts) - Compositor multi-GPU model. Mutter picks one primary GPU, composites there, and copies buffers to displays on other GPUs [SOURCED https://github.com/GNOME/mutter/blob/main/doc/multi-gpu.md]. Any DRM device it can open may still get probed/initialised, which is the likely source of a few MiB of gnome-shell VRAM on a second GPU [INFERRED]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#core-concepts) - PRIME render offload = per-application choice of render GPU while display stays on the primary; pure compute = no graphics API at all. They are different mechanisms; CUDA does not need PRIME [INFERRED from NVIDIA PRIME README scope]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#core-concepts) - Selection is by identifiers, not by order. Device ordering (nvidia-smi index, card0/card1) is not stable across a late-arriving eGPU; use PCI address or UUID [SOURCED CUDA docs https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/environment-variables.html for UUID/PCI_BUS_ID]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#core-concepts) - VGA arbiter / boot_vga. The kernel marks one VGA-class device boot_vga (sysfs /sys/bus/pci/devices//boot_vga); the firmware-initialised device normally wins, so an iGPU usually keeps it when an eGPU appears later. Effects of a late-arriving eGPU on vgaarb are [INFERRED]; the vga_switcheroo kernel page does not cover boot_vga [SOURCED https://docs.kernel.org/gpu/vga-switcheroo.html, negative finding]. An RFC exists to let users select the primary adapter at boot [SOURCED https://lkml.iu.edu/hypermail/linux/kernel/2309.0/02231.html (search result title only)]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#core-concepts) ## xe vs i915 on Arrow Lake - Rule of thumb: Meteor Lake was the last iGPU generation defaulting to i915; Lunar Lake and Battlemage default to xe [SOURCED https://www.phoronix.com/review/intel-mtl-i915-xe-linux via search summary]. Arrow Lake default binding is not stated in fetched sources - check the actual box (below). [INFERRED: Arrow Lake-P follows the xe-default side of that line; unverified] — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#xe-vs-i915-on-arrow-lake) - Both drivers have a force_probe mechanism so exactly one official driver probes a device at a time; .require_force_probe protects platforms still under development [SOURCED https://docs.kernel.org/next/gpu/rfc/xe.html (search snippet)]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#xe-vs-i915-on-arrow-lake) - Parameter syntax (xe): xe.force_probe=[,...], ! before an ID blocks the probe (e.g. 4500,!4571) [SOURCED https://cateee.net/lkddb/web-lkddb/DRM_XE_FORCE_PROBE.html via search snippet]. Switch pattern: i915.force_probe=! xe.force_probe= [SOURCED Phoronix/Xe benchmark, search snippet]. The is the PCI device ID (hex, no 0x): the second half of the [vendor:device] pair that lspci -nn prints, e.g. 8086:xxxx gives xxxx. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#xe-vs-i915-on-arrow-lake) - Kernel Kconfig knobs: CONFIG_DRM_XE_FORCE_PROBE, and the i915 equivalent [SOURCED lkddb link above]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#xe-vs-i915-on-arrow-lake) - [BOX] Both modules loaded and dmesg showed i915 initialised the display: loading a module is not binding. Only the sysfs driver symlink tells the truth. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#xe-vs-i915-on-arrow-lake) - Caveat [INFERRED]: Do not run both drivers against the same device; if you choose xe, block i915 for that ID (i915.force_probe=!) and vice versa. Put options in /etc/modprobe.d/*.conf (options xe force_probe= and options i915 force_probe=!) and regenerate the initramfs (on Ubuntu: sudo update-initramfs -u) so early boot uses them. Blanking hazard [INFERRED]: this iGPU drives the only display, so blocking the driver that currently binds it, when the other driver cannot claim the device, leaves a black screen at next boot. Trial the change once through a one-shot edit of the boot entry (GRUB e, kernel-parameter form) before persisting it in modprobe.d or the default cmdline, change one driver at a time, and keep the old parameters written down. Forum reports of boot hangs after switching on Arrow/Meteor Lake exist [SOURCED https://github.com/strongtz/i915-sriov-dkms/issues/358 (title only)] - keep a rescue entry. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#xe-vs-i915-on-arrow-lake) - Distro note: the kernel 7.0 defaults were not fetched; verify with the commands below rather than assuming. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#xe-vs-i915-on-arrow-lake) ## Primary GPU Selection - Mutter applies a series of rules to select a primary GPU; explicit override is a udev tag mutter-device-preferred-primary (documented example file /etc/udev/rules.d/61-mutter-preferred-primary-gpu.rules, matched by PCI vendor/device attributes; log line on success: GPU /dev/dri/card1 selected primary given udev rule) [SOURCED https://github.com/GNOME/mutter/blob/main/doc/multi-gpu.md]. Older tutorials use ENV{DEVNAME}=="/dev/dri/card0", TAG+="mutter-device-preferred-primary" - do not use it, card numbers shift when the eGPU arrives [SOURCED search result; INFERRED risk]. Match by vendor/device or PCI path, per the mutter doc. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#primary-gpu-selection) - Mutter's doc says nothing about boot_vga in the fetched text; do not rely on it as a documented mutter rule. It is still the practical default outcome when no tag is set [INFERRED]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#primary-gpu-selection) - GNOME Shell extension mutter-primary-gpu exists as a user-facing override [SOURCED https://github.com/zaidka/mutter-primary-gpu]; it is an extension, not the udev tag - do not confuse the names. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#primary-gpu-selection) - To keep the iGPU primary, tag the iGPU explicitly (pin), rather than trying to "untag" NVIDIA. Example shape (verify against the mutter doc before use; it matches on vendor 0x8086 only, no device IDs, so it would also tag any other Intel GPU such as a discrete Arc card): — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#primary-gpu-selection) - [INFERRED from the documented shape; NOT verified end-to-end] A new udev rule applies to cards that appear afterwards; on the live desktop, do not run udevadm trigger against the drm subsystem to force it. Take effect at next boot, then check the log line below. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#primary-gpu-selection) - Check who won: journalctl -b /usr/bin/gnome-shell | grep -i 'selected primary' (message text from the mutter doc; the executable-path match is the more portable form, because the user-unit name for gnome-shell differs between distros and X11/Wayland sessions [INFERRED]). Read-only; if the journal reports no entries for your user, prefix with sudo. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#primary-gpu-selection) - DRM_PRIME here means the kernel PRIME dma-buf sharing between DRM devices (DRIVER_PRIME), which is what lets a secondary GPU's buffers reach the primary; it is not a selection rule [INFERRED]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#primary-gpu-selection) ## Keeping the Desktop Off the eGPU - Scanout: don't plug monitors into the eGPU. That already keeps scanout off it [SOURCED Arch eGPU page]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#keeping-the-desktop-off-the-egpu) - Residual footprint: gnome-shell can still appear in nvidia-smi on a second GPU with a few MiB [BOX 4 MiB]. Known reports show gnome-shell holding VRAM on NVIDIA, growing with extensions/resizing [SOURCED https://bugs.launchpad.net/bugs/1823544 ; https://forums.developer.nvidia.com/t/multiple-wayland-compositors-not-freeing-vram-after-resizing-windows/307939]. On a non-primary GPU the mechanism is [INFERRED]: nvidia-drm exposes a KMS-capable /dev/dri/cardN, mutter opens every DRM device and creates a GL/GBM context on it, which allocates a small context. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#keeping-the-desktop-off-the-egpu) - nvidia-drm.modeset: NVIDIA README (older driver) says DRM KMS is disabled by default and enabled via modeset=1 (modprobe nvidia_drm modeset=1); PRIME needs Linux 3.13+, atomic modeset 4.1+ [SOURCED https://download.nvidia.com/XFree86/Linux-x86_64/580.65.06/README/kms.html - fetch summarised; current defaults differ]. Setting modeset=1 enables the DRIVER_MODESET flag so DRM clients can use modesetting APIs; it does not itself install a framebuffer console [SOURCED search result of forums.developer.nvidia.com "Understanding nvidia-drm.modeset=1"]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#keeping-the-desktop-off-the-egpu) - nvidia-drm.fbdev: from driver 570, fbdev=1 was on by default; when modeset=1 and kernel supports it, nvidia-drm replaces the framebuffer console with a DRM-driven one; disable with fbdev=0 [SOURCED search summary of forums.developer.nvidia.com "Increased idle consumption with driver 570"]. Distros (Arch nvidia-utils) may set both on by default [SOURCED https://wiki.archlinux.org/title/NVIDIA via search]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#keeping-the-desktop-off-the-egpu) - Compute-only option: options nvidia_drm modeset=0 fbdev=0 in modprobe.d. Expected effect: nvidia-drm does not expose a modeset-capable DRM card, so the compositor has nothing to open [INFERRED - NOT verified on driver 610; on newer branches modeset may be default-on or the option may behave differently. Test before trusting]. Alternative: keep nvidia_drm from loading, CUDA still works (needs nvidia, nvidia_uvm) [INFERRED]. A blacklist nvidia_drm line only stops alias autoload; an explicit modprobe nvidia_drm or a dependency pull still loads it, so a hard block needs install nvidia_drm /bin/false [INFERRED]. Never rmmod nvidia_drm on a running desktop to test this; set the option, then load or reboot with the eGPU idle, and read the sysfs parameters and ls /dev/dri (see Verification). Side effects: no Vulkan/EGL-on-DRM offload to the eGPU, no Wayland output on it. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#keeping-the-desktop-off-the-egpu) - [BOX] modules are blocked at boot by a modprobe install ... /bin/false guard and loaded by a service - so put the nvidia_drm options in a /etc/modprobe.d/*.conf file that the loader reads at load time (module options need no initramfs rebuild unless the module is packed into it) and set them before the loading service runs; ordering is covered by the sibling orchestration reference [INFERRED]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#keeping-the-desktop-off-the-egpu) ## PRIME Offload vs Headless Compute - NVIDIA PRIME render offload: set __NV_PRIME_RENDER_OFFLOAD=1 (works for Vulkan and EGL); for GLX also set __GLX_VENDOR_LIBRARY_NAME=nvidia; __VK_LAYER_NV_optimus=NVIDIA_only sorts NVIDIA GPUs first in Vulkan (or non_NVIDIA_only) [SOURCED https://download.nvidia.com/XFree86/Linux-x86_64/580.65.06/README/primerenderoffload.html]. Requires nvidia-drm modeset per that README family [INFERRED from README title context]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#prime-offload-vs-headless-compute) - Mesa: DRI_PRIME=1, or DRI_PRIME=pci-0000_01_00_0 (colons/dots -> underscores) [SOURCED https://wiki.archlinux.org/title/PRIME via search]. Applies to Mesa drivers; NVIDIA proprietary uses the __NV vars. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#prime-offload-vs-headless-compute) - Pure compute: none of these are needed. Only CUDA_VISIBLE_DEVICES (below). Do NOT set __NV_PRIME_RENDER_OFFLOAD in service environments for LLM servers - it does nothing for CUDA and opens graphics contexts [INFERRED]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#prime-offload-vs-headless-compute) ## Decision table - The table is design guidance derived from the sourced sections above; only the rows' cited facts are sourced. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#decision-table) ## Multiple Compute Apps on One GPU - CUDA_VISIBLE_DEVICES accepts integer indices (as ordered by nvidia-smi from 0), GPU UUID strings (as in nvidia-smi -L, abbreviations allowed), and MIG identifiers; listing order changes enumeration order. CUDA_DEVICE_ORDER=FASTEST_FIRST (default) vs PCI_BUS_ID [SOURCED CUDA docs link above]. Because a late-arriving eGPU can change indices, prefer UUID (or set CUDA_DEVICE_ORDER=PCI_BUS_ID). Mismatch: nvidia-smi ordering is PCI-bus-based while CUDA default is fastest-first [INFERRED from the two documented defaults]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#multiple-compute-apps-on-one-gpu) - NVIDIA_VISIBLE_DEVICES is the NVIDIA Container Toolkit/runtime variable (container GPU exposure) - not read by bare-metal CUDA [INFERRED; not fetched this session]. Inside a container, CUDA_VISIBLE_DEVICES then narrows further. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#multiple-compute-apps-on-one-gpu) - Sharing 16 GB: CUDA has no per-process VRAM quota; each runner (Ollama, llama-server, etc.) grabs weights + KV cache; the first to allocate wins and later ones OOM or spill [INFERRED]. [BOX] Ollama runner 4.4 GB + separate LM Studio server + gnome-shell 4 MiB. Budget: sum(weights + KV + ~0.5-1 GB context overhead per process) < total minus headroom. Prefer one shared server over several. Unload idle models (Ollama keep_alive, LM Studio idle TTL) [INFERRED]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#multiple-compute-apps-on-one-gpu) - Time-slicing between processes works by default but no isolation; MPS/MIG are separate mechanisms (MIG unavailable on GeForce) [INFERRED]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#multiple-compute-apps-on-one-gpu) - Track: nvidia-smi --query-compute-apps=pid,process_name,used_memory --format=csv -l 5. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#multiple-compute-apps-on-one-gpu) ## Hazards: eGPU absent or arriving late - Boot without eGPU: the iGPU is boot_vga and the session is unaffected as long as nothing requires the NVIDIA device; services depending on /dev/nvidia* must be conditional [INFERRED]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#hazards-egpu-absent-or-arriving-late) - eGPU arrives after login: a new DRM card node appears; with nvidia_drm modeset on, mutter may hot-add the device (extra probe, possible primary re-evaluation, stutter). With card numbers shifting, any udev rule or script keyed to card0/card1 breaks. Key on PCI address/UUID [INFERRED]. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#hazards-egpu-absent-or-arriving-late) - Display manager restarts: loading nvidia_drm modeset=1 while GDM is up, or a udev rule reload/tag change, can require a session re-login (the mutter-primary-gpu extension notes a re-login is required after switching [SOURCED https://github.com/zaidka/mutter-primary-gpu]). Never systemctl restart gdm to "fix" GPU state on a box where the session matters. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#hazards-egpu-absent-or-arriving-late) - eGPU as primary and it drops: the compositor loses its primary GPU; session ends [INFERRED]. Another reason to keep the desktop on the iGPU. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#hazards-egpu-absent-or-arriving-late) - vgaarb changes: a new VGA-class device may register with the arbiter; boot_vga stays with the firmware device [INFERRED]; check cat /sys/bus/pci/devices/*/boot_vga. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#hazards-egpu-absent-or-arriving-late) - Safe change order [INFERRED]: make one change per boot (kernel driver binding, udev primary tag, nvidia_drm options), verify with the read-only checks after each, and keep a second login path (SSH or a spare TTY) open so a blank display does not leave the box unreachable. modprobe.d and udev edits take effect at the next boot or re-login, not mid-session. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#hazards-egpu-absent-or-arriving-late) - Persistence state / order of modules and unplug are covered in sibling references. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#hazards-egpu-absent-or-arriving-late) ## Anti-patterns - Assuming "module loaded" = "driver bound" for xe/i915 (see checks above). — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#anti-patterns) - Using TAG+="mutter-device-preferred-primary" with DEVNAME=="/dev/dri/card0" (unstable numbering). — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#anti-patterns) - Setting __NV_PRIME_RENDER_OFFLOAD globally or in systemd units for compute services. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#anti-patterns) - Indexing GPUs by integer in CUDA_VISIBLE_DEVICES when the eGPU can appear/disappear. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#anti-patterns) - Blaming gnome-shell's few MiB for OOMs; the real budget is model weights + KV + extra runners. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#anti-patterns) - Restarting GDM or removing nvidia_drm while the desktop is live. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#anti-patterns) - Copying nvidia-drm.modeset=1/fbdev advice from Wayland-on-NVIDIA guides into a compute-only setup. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#anti-patterns) - Trusting forum udev snippets without matching against the mutter doc. — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#anti-patterns) ## Sources - Mutter multi-GPU doc - https://github.com/GNOME/mutter/blob/main/doc/multi-gpu.md (fetched via raw.githubusercontent.com) — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#sources) - NVIDIA README, KMS chapter (driver 580.65.06 path) - https://download.nvidia.com/XFree86/Linux-x86_64/580.65.06/README/kms.html — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#sources) - NVIDIA README, PRIME render offload - https://download.nvidia.com/XFree86/Linux-x86_64/580.65.06/README/primerenderoffload.html — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#sources) - CUDA environment variables - https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/environment-variables.html — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#sources) - Arch Wiki External GPU / PRIME / NVIDIA (via search results; direct fetch blocked) - https://wiki.archlinux.org/title/External_GPU , https://wiki.archlinux.org/title/PRIME , https://wiki.archlinux.org/title/NVIDIA — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#sources) - Kernel docs vga_switcheroo - https://docs.kernel.org/gpu/vga-switcheroo.html ; Xe merge plan - https://docs.kernel.org/next/gpu/rfc/xe.html — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#sources) - Kconfig DRM_XE_FORCE_PROBE - https://cateee.net/lkddb/web-lkddb/DRM_XE_FORCE_PROBE.html — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#sources) - Phoronix MTL i915 vs xe - https://www.phoronix.com/review/intel-mtl-i915-xe-linux (direct fetch 403; summary from search) — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#sources) - NVIDIA developer forum threads on modeset/fbdev - https://forums.developer.nvidia.com/t/understanding-nvidia-drm-modeset-1-nvidia-linux-driver-modesetting/204068 , https://forums.developer.nvidia.com/t/increased-idle-consumption-with-driver-570/321460 — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#sources) - gnome-shell VRAM reports - https://bugs.launchpad.net/bugs/1823544 , https://forums.developer.nvidia.com/t/multiple-wayland-compositors-not-freeing-vram-after-resizing-windows/307939 — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#sources) - mutter-primary-gpu extension - https://github.com/zaidka/mutter-primary-gpu — [source](https://llms-explorer.com/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux/#sources) ## Context files - [Hybrid graphics on Linux: Intel iGPU plus NVIDIA eGPU for headless compute](https://llms-explorer.com/downloads/sources/global-ai-hub/hybrid-graphics-igpu-plus-nvidia-egpu-headless-linux.md)