Hybrid and dual GPU laptops on Omarchy

Omarchy treats a hybrid laptop as an NVIDIA machine plus an integrated GPU. It installs drivers for both cards, Hyprland composites on whichever card owns your panel, and the discrete card idles until something asks for it. The friction is in the details: which card decodes video, which card the browser renders on, and whether the discrete card ever powers down. For single-vendor pages see NVIDIA, Intel GPU and AMD GPU.

Status on 4.0.4

Verified against the v4.0.4 source tree on 2026-09-16. The tracker holds 204 issues matching hybrid or dual GPU terms, 156 of them open.

What works on a stock install: both drivers get installed, the session starts on the integrated card, external monitors wired to either card light up, Vulkan drivers for both vendors are installed, and prime-run from nvidia-utils still offloads a single application to the discrete card.

What is unreliable: video decode routing in browsers, suspend and resume, and discrete GPU runtime power management. Two reporters in #8989 measured zero milliseconds of runtime suspension across a whole session.

The policy is older than Quattro: 3.x wrote the same NVIDIA variables into each user’s config at install time, ungated, and #4901 is the March report of the result. 4.0.0 moved them into nvidia.lua behind a detector call that never returned true, because os.execute inside the Hyprland Lua config could not read an exit status, so 4.0.0 was the one release where hybrid machines were accidentally correct. The 4.0.1 fix to that helper switched the variables back on, which is why users date their video problems to 4.0.1.

What Omarchy does automatically

install/hardware/nvidia.sh splits on GSP firmware. Turing and newer, meaning PCI device IDs at or above 0x1e00, get nvidia-open-dkms, nvidia-utils, lib32-nvidia-utils and libva-nvidia-driver. Maxwell, Pascal and Volta, IDs from 0x1340 up to 0x1e00, get the nvidia-580xx legacy branch. Anything older is refused with a pointer to the Arch wiki. It writes options nvidia_drm modeset=1 to /etc/modprobe.d/nvidia.conf and adds the four NVIDIA modules to /etc/mkinitcpio.conf.d/nvidia.conf for early KMS.

install/hardware/vulkan.sh adds vulkan-intel, vulkan-radeon or vulkan-asahi for whichever integrated card lspci reports, so a hybrid machine gets both vendors’ Vulkan drivers. install/hardware/intel/video-acceleration.sh adds intel-media-driver, libvpl and vpl-gpu-rt for HD, UHD, Xe, Iris, Arc and Panther Lake parts, and libva-intel-driver for GMA.

default/hypr/nvidia.lua sets the per-session environment. The GSP branch exports NVD_BACKEND=direct, LIBVA_DRIVER_NAME=nvidia and __GLX_VENDOR_LIBRARY_NAME=nvidia; the legacy branch exports NVD_BACKEND=egl and __GLX_VENDOR_LIBRARY_NAME=nvidia only. The detectors omarchy-hw-nvidia, omarchy-hw-nvidia-gsp and omarchy-hw-nvidia-without-gsp read cached IDs under /sys/bus/pci/devices instead of running lspci, because reading PCI config space wakes a runtime-suspended card out of D3cold, and that wake by itself takes longer than the 1.5 seconds Hyprland allows for a config reload. That landed in 4.0.0.

omarchy-hw-hybrid-gpu decides whether you have a hybrid setup. It asks supergfxctl -s with a one second timeout and looks for Hybrid in the reply; if supergfxctl is missing or wedged, it falls back to counting VGA, 3D and Display entries in lspci. A wedged supergfxd used to hang the menu, which 4.0.0 fixed.

omarchy-toggle-hybrid-gpu sits behind omarchy toggle hybrid gpu and the Hardware trigger menu (Super + Ctrl + H), documented in the manual chapter on toggles and idle. On first run it installs supergfxctl, writes /etc/supergfxd.conf with mode Hybrid and vfio_enable, and enables the daemon. Switching to Integrated adds a supergfxd drop-in that sleeps five seconds before start, because otherwise the daemon races the display manager and the machine freezes at boot. It also installs the force-igpu sleep hook, which bounces the card Vfio then Integrated after every resume so it powers off again, and switches to Vfio before hibernate so the driver detaches from a card it cannot freeze.

Omarchy never sets AQ_DRM_DEVICES. There is no reference to it in bin/, default/ or config/ at v4.0.4, and omarchy-debug does not collect it either.

Known problems

Issue	Models	Status	Fixed in
#8215 forced LIBVA_DRIVER_NAME=nvidia kills or corrupts browser video	Any iGPU plus Turing or newer NVIDIA. Confirmed on ASUS Vivobook 16, Dell Meteor Lake, HP Arrow Lake, AMD Cezanne plus RTX 3050 Ti	open	not fixed, PR #7851 open
#8989 session-wide NVIDIA env also blocks dGPU runtime suspend	iGPU-primary hybrids, both driver branches	open	not fixed
#4901 Chromium picks NVIDIA EGL, cross-GPU DMA-BUF failures	Intel plus NVIDIA, also AMD plus NVIDIA	workaround	not fixed
#3242 random full freeze from RTD3 PCI reallocation loop	Pascal Optimus laptops, Intel HD 530/630 plus GTX 1060 Mobile	workaround	not fixed
#5274 NVRM assertion storm and crash on lid close under s2idle	Hybrids in supergfxctl Hybrid mode, also dGPU-only laptops	open	not fixed, PR #8546 under test
#4891 compositor freeze on resume, stale frame, dead input	Intel or AMD plus NVIDIA, also dual-NVIDIA desktops	open	not fixed
#10433 nvidia-suspend, resume and hibernate units never enabled	all NVIDIA installs, worst on laptops	open	not fixed
#10350 i915 GPU hang and Hyprland abort when dGPU-driven externals are used	Pascal on the 580xx branch	workaround	not fixed
#8776 by-path AQ_DRM_DEVICES login-loops SDDM with no error	any dual-GPU machine with the variable set	open	not fixed, PRs #8786 and #9063 open
#11868 force-igpu strands the machine in Vfio, toggle cannot recover	Integrated mode, supergfxctl laptops	open	not fixed
#12187 hard freeze about five seconds after login on 4.0.4	one Lenovo RTX 4070 laptop on prebuilt nvidia-open, not DKMS	open	not fixed
#12041 whole-screen rainbow corruption after lid resume	AMD Rembrandt iGPU plus RTX 3050	open	not fixed

The first three rows are one bug seen from three angles. Where the integrated card drives every display, nvidia.lua still exports LIBVA_DRIVER_NAME=nvidia, so libva loads the NVDEC driver on a card that is not compositing. Reporters measured corrupted frames, wallpaper punch-through, or vaEndPicture failed, VA error: internal decoding error in a loop that never completes a frame. Pointing that one variable at the integrated card’s driver, iHD or radeonsi, fixed it for every reporter in the thread who tried it. The NVDEC backend also advertises zero encode entrypoints, so hardware encode is silently lost at the same time.

A related trap from #8989: --disable-features=VaapiVideoDecoder does not fix it. One reporter verified the flag was active in the running process, saw no change, and lost a day to network and compositor theories before finding the environment variable. Disabling the browser’s decoder feature does nothing about the driver name that is still exported session-wide.

Fixes that work

Work in this order.

1. Find out which card owns your panel. Check which cardN has the connected output under /sys/class/drm/. If the NVIDIA card logs Cannot find any crtc or sizes, it has no outputs wired and your machine is muxless. That is the common case, and it is what makes the default VA-API choice wrong here.

2. Fix video decode routing. Add hl.env("LIBVA_DRIVER_NAME", "iHD") to ~/.config/hypr/hyprland.lua, or radeonsi on an AMD integrated card. To skip a relogin also run systemctl --user set-environment LIBVA_DRIVER_NAME=iHD, then restart the browser, since a running process keeps its old environment. This affects encode and decode only, not Vulkan, OpenGL, NVENC or prime-run.

3. If Chromium still ghosts or blacks out, force it onto Mesa EGL per application. A wrapper or desktop file setting __EGL_VENDOR_LIBRARY_FILENAMES=/usr/share/glvnd/egl_vendor.d/50_mesa.json puts Chromium on the same card as the compositor. Set that variable globally and Hyprland’s own EGL init breaks, dropping you to a TTY, so keep it scoped to the browser. See Chromium flicker and hardware acceleration.

4. For suspend problems, start with nvidia-suspend.service, nvidia-resume.service and nvidia-hibernate.service. Omarchy installs the driver but leaves those units disabled, and enabling them together with NVreg_PreserveVideoMemoryAllocations=1 fixed a slow resume in #10433. The evidence is mixed further along: one #5274 reporter crashed with the units already enabled, one #4891 reporter found the sleep script exits immediately on the 610 driver branch, and another #5274 reporter froze on deep sleep, so mem_sleep_default=deep is not a guaranteed fix either. If your machine sleeps with s2idle, the option under test in PR #8546, options nvidia NVreg_EnableS0ixPowerManagement=1 followed by limine-mkinitcpio, has three clean reports in that thread. See suspend will not resume.

5. For random full freezes on older Optimus hardware, pin runtime power management off with options nvidia NVreg_DynamicPowerManagement=0x00 plus a udev rule setting power/control to on for vendor 0x10de, then rebuild the initramfs. Costs battery, stops the PCI reallocation loop.

6. If you never use the discrete card, turn it off. omarchy toggle hybrid gpu switches to Integrated and reboots. Caveat from #11868: if a resume transition fails you can land in Vfio mode, where the toggle prints Hybrid GPU not found or in unknown mode. and exits. The reporter recovered with supergfxctl -m Hybrid, removing /usr/lib/systemd/system-sleep/force-igpu and /etc/systemd/system/supergfxd.service.d/delay-start.conf, then a reboot.

7. Leave AQ_DRM_DEVICES alone unless nothing else works. Never use a /dev/dri/by-path/pci-... name. Aquamarine splits the value on every colon, so the PCI address falls apart, no GPU is found, and the session loops at login with nothing useful logged. Do not pin a bare /dev/dri/cardN either: #8776 saw the numbers rotate between reboots. Use colon-free udev symlink names and list both cards. On btrfs installs the variable lives in ~/.config/uwsm/env-hyprland on the home subvolume, so every snapshot inherits a bad value and rolling back does not help.

8. On 4.0.4, if the machine freezes shortly after login, check which NVIDIA package you have. The release changed the default boot entry to linux-omarchy, and the #12187 reporter traced their freeze to prebuilt nvidia-open, which only ships modules for the stock Arch kernel. That is one report with no follow-up yet. Boot the stock kernel entry from Limine, then switch to nvidia-open-dkms, which is what the installer selects. See kernel panic after update and rollback with Snapper and Limine.

Report it

Run omarchy debug. It gathers inxi -Farz, dmesg, this boot’s journal warnings and errors, and the package list, then offers to upload to logs.omarchy.org with a 24 hour expiry or save locally. Use omarchy debug --print --no-sudo to read it first. If the subcommand is not registered on your install, reported on 4.0.2 and 4.0.3-1 in #11988, call omarchy-debug directly.

Add what it does not capture: which cardN and renderDN map to which PCI address, which card holds the connected outputs, systemctl --user show-environment | grep -E 'LIBVA|GLX|NVD_BACKEND', whether AQ_DRM_DEVICES is set, and /sys/bus/pci/devices/<dGPU>/power/runtime_status. Copy the A/B tables in #8215 and #8989: same clip, same profile, one variable changed. If your laptop is not listed yet, send it to hardware submissions.

Related

• NVIDIA on Omarchy 4, hybrid GPU black screen
• Suspend and sleep, battery and power, multi-monitor
• Battery drains fast, still broken on the current release, v4.0.4 notes