Skip to content

Enable nvidia-powerd so laptop GPUs get Dynamic Boost - #10592

Open
onelegdave wants to merge 8 commits into
omacom:quattrofrom
onelegdave:enable-nvidia-powerd
Open

onelegdave wants to merge 8 commits into
omacom:quattrofrom
onelegdave:enable-nvidia-powerd

Conversation

@onelegdave

@onelegdave onelegdave commented Sep 7, 2026 •

Copy link
Copy Markdown

Fixes #9678

What

Enables nvidia-powerd in install/hardware/nvidia.sh after driver installation, only when an ACPI device exposes NVIDIA's platform-controller hardware ID (NVDA0820).

Why

Arch ships nvidia-powerd.service with preset: disabled. Nothing in the install enables it, so on laptops the GPU stays at its base TGP with no Dynamic Boost.

On an ASUS ROG Strix G16 with an RTX 5060 Laptop GPU:

Without the daemon:

Current Power Limit    : 50.00 W
Default Power Limit    : 50.00 W

Pinned at 49.5–50W at 65–69°C under sustained 99% load.

With it:

Current Power Limit    : 100.00 W
Default Power Limit    :  50.00 W
Max Power Limit        : 115.00 W

85–93W at 78–80°C under the same load.

Reproduced in both directions on the same hardware — masking the daemon on a working install drops the ceiling straight back to 50W, and enabling it on a fresh Omarchy install raises it immediately.

Notes

  • enable rather than enable --now, since the install runs before the graphical session and the daemon is only useful from next boot.
  • Checks for an exact NVDA0820 match in /sys/bus/acpi/devices/*/hid, without assuming instance :00. Systems without a match skip enablement.
  • Unsupported systems can leave a failed service because the daemon can exit nonzero and the unit restarts on failure. The guard avoids enabling it on systems without the controller; the controller's presence does not guarantee working Dynamic Boost.
  • Discussion Acer gaming laptop GPU power management setup #4726 reports the same symptom on Acer laptops.

Arch ships nvidia-powerd.service with preset: disabled. Without it,
laptop NVIDIA GPUs stay at base TGP with no Dynamic Boost.

On an RTX 5060 Laptop this meant a hard 50W ceiling against 100W with
the daemon running - 49.7W sustained versus 85-93W under identical load.
Reproduced in both directions on the same hardware.

Harmless on desktop cards, where the daemon finds nothing to manage.
@kh4rit-bot

Copy link
Copy Markdown

Confirming on a third vendor. Lenovo Legion Pro 5 16ARX8, Ryzen 7 7745HX, RTX 4060 Laptop GPU, Omarchy 4.0.3, nvidia-open-dkms 615.71.09, kernel 7.2.3.

Fresh Omarchy install had nvidia-powerd.service disabled and the GPU capped at exactly its default:

Current Power Limit : 55.00 W
Default Power Limit : 55.00 W
Max Power Limit     : 140.00 W

Under a sustained VR load (SteamVR + Half-Life: Alyx) it sat at 97-99 % utilization, throttling from 2220 MHz to about 1600 MHz.

After systemctl enable --now nvidia-powerd the ceiling went to 140 W immediately, without a reboot. The same load then held 2100-2280 MHz. The daemon logged only "DBus Connection is established", no NvPCF errors, even though the driver prints the usual PlatformRequestHandler failed to get platform power mode from SBIOS assertions at boot on this machine. Those did not block Dynamic Boost.

One note on the "harmless on desktop cards" point. The unit ships with Restart=on-failure and StartLimitBurst=5. On systems without Dynamic Boost support the daemon exits non-zero (NVIDIA forum reports show "Failed to initialize RM Client" followed by Failed with result 'exit-code'), so it would retry five times and leave a failed unit on every desktop. Gating on the ACPI device Dynamic Boost requires would avoid that, for example: [[ -e /sys/bus/acpi/devices/NVDA0820:00 ]] && systemctl enable nvidia-powerd. That device (\_SB_.NPCF, HID NVDA0820) is present on this Legion and is what the daemon talks to.

@onelegdave

Copy link
Copy Markdown
Author

Thanks for confirming this on the Lenovo and for catching the desktop failure behavior. I've added the ACPI guard in a3633d9 and corrected the description's "harmless on desktop cards" claim.

The installer now checks for an exact NVDA0820 hardware ID in /sys/bus/acpi/devices/*/hid, so it does not depend on the device being instance :00. Systems without a match skip enablement. It still uses systemctl enable, leaving startup to the next boot.

Bash syntax, whitespace checks, and the existing NVIDIA hardware-detection and KMS-hook tests passed. Eight mocked installer scenarios also passed, covering absent/unrelated devices, instances :00 and :01, a similar nonmatching ID, multiple matches, no NVIDIA GPU, and service-enabling failure. Successful NVIDIA paths still write the modprobe and mkinitcpio configuration.

This detects the platform-controller device; it does not guarantee working Dynamic Boost on every firmware/driver combination. The guard revision has not been tested across other physical laptops or in a fresh ISO installation.

@omarchybot omarchybot added bug Something isn't working compatibility Make Omarchy work better on everything! labels Sep 27, 2026
@omarchybot

Copy link
Copy Markdown
Collaborator

Reviewed at a3633d9 against quattro (821ae58). Nothing wrong found in the change.

What was checked, on a disposable Omarchy VM:

  • ./test/cli passed, as did hw-nvidia-test.sh, nvidia-kms-hook-test.sh and hw-display-test.sh.
  • The guard was run against fake ACPI trees with lspci, omarchy-pkg-add and systemctl stubbed. It enables nvidia-powerd for NVDA0820 at :00 and at :01, and skips a near-miss ID (NVDA08201), unrelated devices, and an empty directory. The base branch never enables it in any of those cases. Both branches still write the modprobe and mkinitcpio files.
  • On the VM's real ACPI tree, which has no NVDA0820, the grep returns 1 quietly. It sits inside the if, so it can't abort setup under set -e.
  • Both packages the script installs ship the unit: usr/lib/systemd/system/nvidia-powerd.service is in nvidia-utils 610.57.04 and in nvidia-580xx-utils 580.178.04. So systemctl enable can't fail on a missing unit in either branch.
  • Failing to boot into the graphical desktop environment after enabling nvidia-powerd.service #3079 reported a laptop that would not reach the desktop after enabling nvidia-powerd. That reporter fixed it with early KMS, which this script already writes (nvidia_drm modeset=1, and the nvidia modules in mkinitcpio).

A second opinion from Codex Medium read the diff and the base branch's rules and found no defect. It agreed on the guard, the unit being in both packages, and #3079. It also notes that NVDA0820 shows the platform controller is present but does not prove Dynamic Boost works on that firmware. Its independence from this review is not guaranteed.

Two things are left, and both belong to the maintainer:

  • The power-limit change itself only shows up on an NVIDIA laptop, and a VM can't show it. The evidence so far is the measurements in Laptop NVIDIA GPUs stuck at default TGP — nvidia-powerd is never enabled #9678 and in this PR, from the account that opened both, plus the Legion report above, which was taken before the guard was added. So this needs confirming on laptop hardware at this commit before it is marked verified.
  • Only new installs get this. There is no migration, so existing laptops stay at their default TGP until nvidia-powerd is enabled by hand. Whether to add one is a rollout decision.

Waiting on the maintainer.

@szoniu

szoniu commented Oct 4, 2026

Copy link
Copy Markdown

Confirming on an Intel hybrid laptop: ASUS ROG Zephyrus G16 GU605MI (Core Ultra 9 185H, RTX 4070 Laptop GPU), Omarchy 4.0.4, kernel 7.2.5-3-omarchy, nvidia-open-dkms 610.57.04. MUX in Hybrid mode, so the Intel iGPU drives the internal panel and the NVIDIA GPU renders through PRIME. NVDA0820:00 is present, so the guard in a3633d9 matches this machine.

On a fresh install (2026-10-03) nvidia-powerd.service was disabled and the GPU sat at its default limit:

Current Power Limit : 55.00 W
Default Power Limit : 55.00 W
Max Power Limit     : 115.00 W

Under a game load (Overwatch through DXVK) it was at 100% utilization, ~55 W and 1450–1650 MHz, with clocks_event_reasons = 0x4 (SW power cap).

After systemctl enable --now nvidia-powerd the current limit went to 85 W right away, and to about 104 W at idle after a reboot, so the service comes up on boot. In the same game the limit moves between ~75 and ~101 W as Dynamic Boost shifts power between CPU and GPU; the GPU runs at 1900–2500 MHz and 75–100 W until it reaches the 87 °C temperature target.

I enabled the service by hand rather than running the PR's install script.

One thing I couldn't check here: whether the daemon keeps the dGPU from runtime-suspending. On this machine the dGPU never suspends anyway, because Firefox's media process (RDD) holds it, presumably through LIBVA_DRIVER_NAME=nvidia (#11173), so I can't separate the two.

@sergihrs

sergihrs commented Oct 4, 2026

Copy link
Copy Markdown

Tested the change in a3633d9 on an ASUS ROG Zephyrus G16 GU605CP (DMI GU605CP_GU605CP, BIOS 307):
Core Ultra 9 285H + RTX 5070 Laptop 10de:2d58 (Blackwell) in hybrid mode, Omarchy 4.0.4, kernel
7.2.5-3-omarchy, nvidia-open-dkms 610.57.04. The guard matches (NVDA0820:00). I enabled the unit by hand
(same end state as the installer step) plus a local ConditionACPower=true drop-in, which changes nothing on AC.

  • Before (unit disabled, on AC): Current Power Limit 55 W = Default, Max 115 W. Limits only; I didn't run a
    load then.
  • After (unit enabled, on AC): Current Power Limit 105 W, the spec 90 W TGP + 15 W Dynamic Boost. A 50 s CUDA
    load (ffmpeg bilateral_cuda) averaged 104.5 W (101–107 W) at 2385–2475 MHz, 85 °C max, with SW power cap
    active throughout, so the 105 W limit is what binds.
  • Runtime suspend with the daemon running (the open question in the GU605MI report): the dGPU stayed suspended
    for a 60 s idle window (0 ms active) and suspended again 18 s after the load.

On ROG, asusd also starts nvidia-powerd on AC plug-in, but not at boot, so booting on AC stays at 55 W without
this PR. This machine was updated to 4.0.4 and the unit stayed disabled, which matches the migration point above.

@omarchybot

Copy link
Copy Markdown
Collaborator

Two new hardware reports since the last review, read at a3633d9. The head hasn't moved and install/hardware/nvidia.sh is unchanged on quattro, so the earlier results still stand: the suite passed on a disposable Omarchy VM and the guard behaved correctly against fake ACPI trees.

What the new reports add:

  • Both laptops expose NVDA0820:00, so the guard matches them: an ASUS Zephyrus G16 GU605MI (RTX 4070) and a GU605CP (RTX 5070). On both, enabling nvidia-powerd raised the power limit from the 55 W default (to 85–105 W), and it held under sustained load. On the GU605MI the service also came up after a reboot.
  • The GU605CP report answers the runtime-suspend question from the GU605MI report: with the daemon running, the dGPU stayed suspended at idle and suspended again after load. That's one machine.
  • It also shows why this belongs at install: on ROG laptops asusd starts the daemon when AC is plugged in, but not at boot.

Neither reporter ran this PR's install script. Both enabled the unit by hand, which ends in the same service state but doesn't exercise the guard. The GU605CP also added a ConditionACPower=true drop-in. That makes no difference on AC, but it means nobody has tested the PR's unconditional enable on battery yet.

A second opinion from Codex Medium read both reports and came to the same conclusions. It also raised the battery point, which this review had not. It found no new defect. Its independence from this review is not guaranteed.

This still isn't marked verified. Hardware confirmation has to come from the issue's reporter (here the PR's author, who can't confirm their own fix) or a maintainer, naming the full commit. The open questions are unchanged: confirming on a laptop at this commit, possibly including battery behaviour, and whether existing installs should get a migration. The GU605CP report shows an updated machine still has the unit disabled.

Waiting on the maintainer.

@szoniu

szoniu commented Oct 19, 2026

Copy link
Copy Markdown

Battery test on the same ASUS Zephyrus G16 GU605MI (Core Ultra 9 185H + RTX 4070 Laptop, hybrid mode), at a3633d9. Same caveat as before: the unit is enabled by hand (systemctl enable nvidia-powerd), not by the PR's script, so this checks the end state, not the guard. Driver 615.78.08 (nvidia-open), kernel 7.2.8-9-omarchy, asusctl 6.5.0.

Booted on battery: the enabled unit came up at boot and logged ERROR! Client (presumably SBIOS) has requested to disable Dynamic Boost DC controller, which looks like the firmware switching Dynamic Boost off on DC rather than a fault. It kept running, no restarts.

Idle on battery, 2 min settle + 60 s measured, screen on at fixed brightness (idle inhibited), no game or browser on the dGPU:

dGPU suspended battery draw Current Power Limit
nvidia-powerd running 60/60 s 12.6 W 47 W
nvidia-powerd stopped 58/60 s 12.2 W 42 W

(Default 55 W, Max 115 W in both.) So with the daemon running on battery the dGPU still runtime-suspends at idle, and the 0.4 W difference is within noise.

AC transitions with the daemon running: plugging in raised Current Power Limit 47 -> 105 W; unplugging dropped it to 42 W because asusd stopped the service. disable_nvidia_powerd_on_battery: true is the default in /etc/asusd/asusd.ron (asusd logs External power supply state changed: false, then systemd Stopping nvidia-powerd service...). So on ROG laptops asusd already manages the daemon on battery; the unconditional enable only matters until the first AC -> battery switch after boot, and it caused no problem in that window.

omarchybot and others added 2 commits October 21, 2026 14:33
The GB10 desktops (DGX Spark, Ascent GX10) reach this branch and expose the same NVDA0820 platform controller, but NVIDIA documents Dynamic Boost for Intel and AMD notebooks only, so the guard alone would enable the daemon on an unsupported desktop.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Codex Medium <noreply@openai.com>
@omarchybot

Copy link
Copy Markdown
Collaborator

Re-reviewed because quattro changed install/hardware/nvidia.sh in #14820 (NVIDIA N1x and aarch64 support) and a new battery report came in. I pushed two commits to this branch. The head is now 2c6f70a.

  • 1491d9c merges quattro into the branch. It merges cleanly. The guard still sits after omarchy-pkg-add in the generic NVIDIA branch.
  • 2c6f70a makes the guard omarchy-hw-x86 && grep -qx 'NVDA0820' .... The GB10 desktops (DGX Spark, ASUS Ascent GX10) are an Omarchy install target and take this same branch, and they expose NVDA0820:00 too: a published DGX Spark hardware inventory lists it. NVIDIA's Dynamic Boost README supports it only on Intel or AMD notebooks. Without the x86 check, a GB10 install would enable the daemon on a desktop it doesn't support, and avoiding that is what the guard is for. The merge is there because omarchy-hw-x86 first shipped in Support the NVIDIA N1x laptops #14820.

What was checked, on a disposable Omarchy VM at 2c6f70a:

  • ./test/cli passed, and so did hw-nvidia-test.sh, nvidia-kms-hook-test.sh and nvidia-install-packages-test.sh.
  • I ran the script against fake ACPI trees with lspci, omarchy-pkg-add and systemctl stubbed. On x86 it enables nvidia-powerd for NVDA0820 at :00 and at :01. It skips a near-miss ID, unrelated devices and an empty tree. With NVDA0820 present it skips on aarch64, and an N1x still takes its own branch. Every non-N1x case still writes the modprobe and mkinitcpio files. With the guard removed, nothing is enabled in any case.

The battery report on the GU605MI answers the question raised last time, for that machine: the daemon came up on battery, kept running after the firmware turned off the DC controller, and the dGPU still runtime-suspended at idle. On ROG laptops asusd stops the service on unplug anyway. This was tested on one ASUS model with the unit enabled by hand, not through this script.

Codex Medium found the GB10 case in its review of the merged tree. It then reviewed the fix and found no defect.

This still isn't marked verified. A Dynamic Boost fix can only be confirmed on a laptop, and it needs a maintainer, testing the full commit 2c6f70a. Whether existing installs should get a migration is still open. Waiting on the maintainer.

@greptile-apps

greptile-apps Bot commented Oct 21, 2026 •

Copy link
Copy Markdown
Contributor

RetriggerView in GreptileConfidence Score: 5/5

[Medium impact] The PR appears safe to merge.

Summary

Enables nvidia-powerd after NVIDIA driver installation when an x86 machine exposes the exact NVDA0820 hardware ID.

  • NVIDIA setup enables nvidia-powerd on matching x86 systems.

Reviews (3) · Last reviewed commit: "Reset the sandboxed /etc between test ca..." · Reviewed by Greptile

Comment thread install/hardware/nvidia.sh Resolved
onelegdave and others added 4 commits October 26, 2026 12:02
The section says Codex helped with tests "in this fork", which stops being true once it lands in Omarchy's README, and it is unrelated to enabling nvidia-powerd. The PR description is where that disclosure belongs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
AGENTS.md requires the exact platform names in tests; `arm` is not one of them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Muse Wonder High <noreply@omarchy.org>
The files the first case wrote let every later case pass the "completes setup after the guard" check without the script reaching those writes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Codex Medium <noreply@openai.com>
Co-Authored-By: Muse Wonder High <noreply@omarchy.org>
@omarchybot

Copy link
Copy Markdown
Collaborator

Re-reviewed because the branch gained 6619015, which adds test/shell.d/nvidia-powerd-test.sh and a README section. install/hardware/nvidia.sh is the same as at the last review, and quattro hasn't changed it since. I pushed three commits on top, so the head is now e6c48d7.

The new test runs the whole setup script against a fake ACPI tree and a sandboxed /etc. On a disposable Omarchy VM it passes at this head. It fails against quattro's nvidia.sh (x86 with NVDA0820 enables nothing) and fails with the omarchy-hw-x86 check removed (the aarch64 case enables the daemon), so it guards both halves of the condition. ./test/cli, hw-nvidia-test.sh, nvidia-kms-hook-test.sh, nvidia-install-packages-test.sh and nvidia-apps-test.sh also pass.

What I pushed, @onelegdave:

  • 883b6cd removes the README "AI assistance" section. It says Codex helped "in this fork", which stops being true once it's in Omarchy's README, and it has nothing to do with enabling nvidia-powerd. The PR description is the place for that note. If you meant it for upstream, it would need its own PR.
  • 485739d changes the test's arm platform token to aarch64, because AGENTS.md asks for the exact platform names in tests.
  • e6c48d7 clears the sandboxed /etc before each case. Without that, the files the first case wrote let every later case pass the "completes setup after the guard" check. With the old test, a script that exits right after the guard on a negative case still passed 7/7. With this change it fails.

Codex Medium and muse-wonder (via pi, thinking high) reviewed the new commits. Both found the stale /etc problem, and muse-wonder found the platform name. Both then reviewed the three fixes and found nothing left. Their independence from this review is not guaranteed.

This still isn't marked verified. Dynamic Boost only shows up on an NVIDIA laptop, and #9678 was opened by the author of this PR, so the confirmation has to come from a maintainer, testing the full commit e6c48d7. Whether existing installs should get a migration is also still open. Waiting on the maintainer.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

bug Something isn't working compatibility Make Omarchy work better on everything!

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Laptop NVIDIA GPUs stuck at default TGP — nvidia-powerd is never enabled

5 participants