Nvidia RTX Laptop Linux Power Management

Get proper battery life on Linux laptops with NVIDIA RTX GPUs.

TL;DR: By default, NVIDIA GPUs stay awake and drain ~6W constantly. This guide enables D3cold power states so the GPU sleeps when idle, doubling battery life on hybrid graphics laptops (tested: 20W → 10W idle).

The Problem

Linux hybrid graphics laptops with NVIDIA RTX GPUs have terrible battery life out of the box. The GPU never enters deep sleep (D3cold) because:

1. GNOME Shell uses NVIDIA for EGL rendering - libglvnd loads 10_nvidia.json before 50_mesa.json, so the desktop compositor keeps the GPU awake permanently.

2. The NVIDIA driver has bugs - Sleep/shutdown hangs, CPU spin bugs, and missing power management features require workarounds.

3. System services keep the GPU awake - nvidia-persistenced, nvidia-settings, and even Chrome probe the GPU at startup.

4. TLP fights against you - Default RUNTIME_PM_ON_AC=on forces all devices to stay awake on AC power.

Quick Start

git clone https://github.com/jvdillon/rtx-laptop-linux.git
cd rtx-laptop-linux
sudo ./setup.sh
# Reboot required

The script will:

1. Install NVIDIA drivers (or prompt to reinstall if present)
2. Apply all power management fixes
3. Configure systemd services for proper sleep/wake handling

What the Setup Does

1. Demote NVIDIA EGL Priority

libglvnd loads EGL vendors by filename order; 10_nvidia.json wins over 50_mesa.json, so the desktop renders on (and wakes) the dGPU. A plain rename is not durable: libnvidia-gl upgrades re-create 10_nvidia.json, silently undoing the demotion. Use dpkg-divert so the package's own file is permanently redirected and never reappears at the winning filename across upgrades.

sudo dpkg-divert --add --rename \
    --divert /usr/share/glvnd/egl_vendor.d/90_nvidia.json \
    /usr/share/glvnd/egl_vendor.d/10_nvidia.json

2. Force Mesa EGL System-Wide

Even with the priority fix, some apps probe NVIDIA. This environment variable forces Mesa for all desktop rendering:

# /etc/environment
__EGL_VENDOR_LIBRARY_FILENAMES=/usr/share/glvnd/egl_vendor.d/50_mesa.json

3. Disable NVIDIA DRM Modesetting

The dGPU drives no display connectors (all eDP/DP/HDMI are on the Intel iGPU), so KMS only exposes /dev/dri/card0+renderD129 for the desktop to probe. CUDA uses /dev/nvidia* directly and is unaffected by modeset.

Gotcha: ubuntu-drivers ships /usr/lib/modprobe.d/nvidia-kms.conf with modeset=1. modprobe merges all *.conf across /etc and /usr/lib, sorts by filename, and for a repeated option last-wins. A differently-named file in /etc does not override it — nvidia-graphics-drivers-kms.conf ("g") sorts before nvidia-kms.conf ("k"), so the /usr/lib modeset=1 wins. The only deterministic override is a same-basename file in /etc, which fully shadows the /usr/lib copy.

# Same basename as /usr/lib/modprobe.d/nvidia-kms.conf -> /etc copy replaces it
echo 'options nvidia-drm modeset=0' | sudo tee /etc/modprobe.d/nvidia-kms.conf
sudo update-initramfs -u

Verify after reboot: cat /sys/module/nvidia_drm/parameters/modeset -> N.

4. Fix nv_open_q CPU Spin Bug

Driver bug: the nv_open_q kernel thread spins at 20-70% CPU when nonblocking open is enabled (default).

# /etc/modprobe.d/nvidia-graphics-drivers.conf
options nvidia NVreg_EnableNonblockingOpen=0

Reference: NVIDIA/open-gpu-kernel-modules#615

5. Enable Runtime PM via udev

The GPU needs power/control=auto to enter D3cold. This udev rule sets it for all NVIDIA 3D controllers:

# /etc/udev/rules.d/80-nvidia-pm.rules
ACTION=="add|change|bind", SUBSYSTEM=="pci", ATTR{vendor}=="0x10de", ATTR{class}=="0x030200", ATTR{power/control}="auto"

6. Wake GPU Before Shutdown/Sleep

Driver bug: when the GPU is in D3cold, the driver can't cleanly unload or prepare for system sleep. This causes hangs after "Reached target Reboot" or sleep failures with error -5.

A systemd service wakes the GPU before any shutdown/sleep target:

# /etc/systemd/system/nvidia-wake.service
[Unit]
Description=Wake NVIDIA GPU before shutdown/sleep to prevent hangs
DefaultDependencies=no
Before=shutdown.target reboot.target halt.target suspend.target hibernate.target hybrid-sleep.target

[Service]
Type=oneshot
ExecStart=/bin/bash -c 'echo on > /sys/bus/pci/devices/0000:01:00.0/power/control; sleep 1'

[Install]
WantedBy=halt.target reboot.target shutdown.target suspend.target hibernate.target hybrid-sleep.target

7. Fix Hibernate Resume

Driver bug: when nvidia modules are in initramfs, they load BEFORE hibernate resume. The kernel then tries to freeze a freshly-loaded driver that was never properly initialized, causing pci_pm_freeze(): nv_pmops_freeze returns -5.

Fix: exclude nvidia from initramfs so modules load AFTER hibernate resume.

Ubuntu 25.04+ (dracut):

# /etc/dracut.conf.d/nvidia-exclude.conf
omit_drivers+=" nvidia nvidia-drm nvidia-modeset nvidia-uvm "

Ubuntu 24.04 LTS (initramfs-tools):

# /etc/initramfs-tools/hooks/exclude-nvidia
# Hook script that removes nvidia modules from initramfs

Also adds early i915 KMS for external display during LUKS decrypt.

References:

• https://bbs.archlinux.org/viewtopic.php?id=285508
• https://forums.developer.nvidia.com/t/preservevideomemoryallocations-systemd-services-causes-resume-from-hibernate-to-fail/233643

8. Restore Runtime PM After Boot/Resume

The nvidia driver sets power/control=on when it initializes, and our wake service sets it before sleep. Both prevent D3cold after boot/resume.

Two systemd units restore power/control=auto:

# /etc/systemd/system/nvidia-power-control.service (after boot)
[Unit]
Description=Enable NVIDIA GPU runtime power management
After=multi-user.target tlp.service

[Service]
Type=oneshot
ExecStart=/bin/bash -c 'echo auto > /sys/bus/pci/devices/0000:01:00.0/power/control'

[Install]
WantedBy=multi-user.target

# /etc/systemd/system/nvidia-resume.service.d/restore-pm.conf (after resume)
[Service]
ExecStartPost=/bin/bash -c 'echo auto > /sys/bus/pci/devices/0000:01:00.0/power/control'

9. Disable nvidia-persistenced

This service keeps the driver loaded even when the GPU is unused, preventing D3cold. Disabling it enables on-demand loading.

Trade-off: first CUDA call has ~1s latency while the driver loads.

sudo systemctl disable nvidia-persistenced

9b. Disable nvidia-powerd (Dynamic Boost)

nvidia-powerd is the Dynamic Boost daemon: while the GPU is active it rebalances the shared CPU/GPU power budget. It depends on the SBIOS exposing the GPU power rail via the NVPCF ACPI interface. This laptop's SBIOS does not, and actively requests Dynamic Boost be disabled — powerd logs:

ERROR! Client (presumably SBIOS) has requested to disable Dynamic Boost DC controller

With powerd running, it polls this dead interface and floods the journal every ~22s with PRH failed to update thermal limit. Dynamic Boost is useless for sustained compute (the GPU wants its full budget) and unavailable here regardless.

Masking powerd removes the every-22s flood. It does not eliminate PRH messages entirely: the driver itself probes the SBIOS thermal interface at init and on each D3cold wake, so a few PRH lines per boot remain (clustered at boot/wake, not the periodic flood). These are benign — there is no driver flag to disable the PlatformRequestHandler (modinfo nvidia exposes none), and the real fix would be an LG BIOS that exposes NVPCF (none exists for the 16Z90TR as of this writing). D3cold/battery and suspend use separate subsystems and are unaffected.

Use mask (not disable): the driver package's systemd preset re-enables powerd on every upgrade; mask survives presets.

sudo systemctl mask --now nvidia-powerd

10. Enable nvidia suspend/hibernate/resume Services

These official NVIDIA services handle saving/restoring GPU state across suspend and hibernate:

sudo systemctl enable nvidia-suspend nvidia-hibernate nvidia-resume

11. Disable nvidia-settings Autostart

nvidia-settings spawns at login and keeps /dev/nvidia* open, preventing D3cold.

# ~/.config/autostart/nvidia-settings-autostart.desktop
[Desktop Entry]
Type=Application
Hidden=true

12. Prevent Chrome from Waking GPU

Chrome probes /dev/nvidiactl at startup even when not using GPU acceleration. Force it to use Mesa:

sudo sed -i 's|^Exec=/usr/bin/google-chrome-stable|Exec=env __NV_PRIME_RENDER_OFFLOAD=0 __GLX_VENDOR_LIBRARY_NAME=mesa /usr/bin/google-chrome-stable|' \
    /usr/share/applications/google-chrome.desktop

13. Configure TLP (if installed)

TLP defaults to RUNTIME_PM_ON_AC=on which forces all devices awake on AC power:

sudo sed -i '/^#*RUNTIME_PM_ON_AC/d; $ a RUNTIME_PM_ON_AC="auto"' /etc/tlp.conf

14. Rebuild initramfs

Apply the initramfs changes (nvidia exclusion, early i915):

sudo update-initramfs -u  # Ubuntu 24.04
sudo dracut --force       # Ubuntu 25.04+

Verification

After rebooting, verify the GPU enters D3cold when idle:

# Should show "suspended" when GPU is idle
cat /sys/bus/pci/devices/0000:01:00.0/power/runtime_status

# nv_open_q should be ~0% CPU (or not running)
ps aux | grep '\[nv_open_q\]'

# Power draw should be 8-12W idle (without heavy apps)
cat /sys/class/power_supply/BAT*/power_now | awk '{print $1/1000000 " W"}'

To wake the GPU for CUDA work:

nvidia-smi  # GPU wakes on-demand

Troubleshooting

GPU won't enter suspended state

Check what's keeping it awake:

# List processes using nvidia devices
sudo fuser -v /dev/nvidia*

# Check if any GL apps are running
lsof /dev/dri/card*

Common culprits: Discord, Slack, VS Code (Electron apps), OBS.

System hangs on shutdown/reboot

The wake service may not be running:

sudo systemctl status nvidia-wake.service
sudo systemctl enable nvidia-wake.service

Hibernate resume fails

Check if nvidia modules are in initramfs:

lsinitramfs /boot/initrd.img-$(uname -r) | grep nvidia

If they appear, the dracut config didn't take effect:

sudo update-initramfs -u

Screen corruption after resume

Try adding to kernel command line:

nvidia.NVreg_PreserveVideoMemoryAllocations=1

Hybrid-sleep doesn't work

Known NVIDIA driver bug (nv_restore_user_channels fails). Use suspend or hibernate instead - hybrid-sleep is fundamentally broken with these drivers.

Finding Your GPU PCI Address

The setup script auto-detects your GPU PCI address. The examples in this README use 0000:01:00.0 for illustration.

To find yours manually:

lspci -D | grep -i nvidia
# Example output: 0000:01:00.0 3D controller: NVIDIA Corporation...

If you need to update the systemd services manually, edit:

• /etc/systemd/system/nvidia-wake.service
• /etc/systemd/system/nvidia-power-control.service
• /etc/systemd/system/nvidia-resume.service.d/restore-pm.conf

Driver Installation (Ubuntu)

For reference, here's how to install the NVIDIA driver and CUDA toolkit on Ubuntu 24.04+:

# Remove any existing nvidia packages
sudo apt purge -y '^nvidia-.*' '^libnvidia-.*' '^cuda-.*' '^libcuda-.*'
sudo apt autoremove -y

# Add NVIDIA CUDA repository
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb
sudo apt update

# Install open driver + CUDA toolkit
sudo apt install -y nvidia-open cuda-toolkit

# Add CUDA to PATH (~/.bashrc)
export PATH="/usr/local/cuda/bin${PATH:+:${PATH}}"
export LD_LIBRARY_PATH="/usr/local/cuda/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}"

Enabling Profiling

To use nsys and other NVIDIA profiling tools without root:

# Allow profiling for all users
echo "options nvidia NVreg_RestrictProfilingToAdminUsers=0" | \
    sudo tee /etc/modprobe.d/nvidia-profiling.conf
sudo update-initramfs -u

# Set perf_event_paranoid
echo "kernel.perf_event_paranoid=2" | sudo tee /etc/sysctl.d/99-perf.conf
sudo sysctl -p /etc/sysctl.d/99-perf.conf

References

• NVIDIA Dynamic Power Management
• Arch Wiki: NVIDIA Tips
• Kernel PCI Power Management
• nv_open_q CPU spin bug

Tested On

• LG Gram Pro 16" 2025 (RTX 5050 Mobile) - Ubuntu 25.10, 25.04
• Ubuntu 24.04 LTS supported (uses initramfs-tools instead of dracut)
• Should work on any hybrid graphics laptop with RTX 20-series or newer

License

Apache 2.0