nvidia-suspend-fix

Fix NVIDIA GPU communication loss after suspend/resume on Linux laptops.

Problem

After suspend or screen lock, nvidia-smi fails and the GPU becomes unresponsive — requiring a full reboot. This is a well-known issue on hybrid GPU laptops (NVIDIA + Intel/AMD iGPU) running recent NVIDIA drivers on Linux.

Symptom: nvidia-smi returns "Unable to communicate with NVIDIA driver" after resume.

Root cause: The kernel module parameter NVreg_PreserveVideoMemoryAllocations is not set by default. Without it, the NVIDIA systemd suspend/resume services run but do nothing — VRAM is not saved/restored across sleep cycles. The kernel log confirms this with Xid Error 31 (MMU Fault, PDE access violation) on resume.

Solution (Fix 1 — works for most people)

sudo ./fixes/01_preserve_vram_apply.sh
sudo reboot

This creates /etc/modprobe.d/nvidia-power-management.conf with:

options nvidia NVreg_PreserveVideoMemoryAllocations=1
options nvidia_drm fbdev=1

And enables the NVIDIA systemd suspend/hibernate services.

Verify after reboot:

cat /sys/module/nvidia/parameters/PreserveVideoMemoryAllocations  # should print 1

Revert

sudo ./fixes/01_preserve_vram_revert.sh
sudo reboot

Additional Fixes

If Fix 1 alone doesn't solve it, apply incrementally:

#	Script	What it does	Confidence	Reboot?
1	fixes/01_preserve_vram_*	Enable VRAM preservation + fbdev	High	Yes
2	fixes/02_persistence_mode_*	Enable nvidia-persistenced persistence mode	Medium	No
3	fixes/03_pcie_aspm_*	Disable PCIe ASPM via GRUB	Low	Yes

Each fix has an _apply.sh and _revert.sh script.

Diagnostic Tools

diagnose.sh — GPU state snapshot

Captures a full snapshot: nvidia-smi output, loaded modules, kernel parameters, PCIe link status, Xid errors, power state, and systemd service status.

sudo ./diagnose.sh baseline        # before any changes
sudo ./diagnose.sh post_fix1       # after applying fix 1
sudo ./diagnose.sh post_resume     # after a suspend/resume cycle

Output is saved to diagnostics/<label>_<timestamp>.log.

test_suspend.sh — automated suspend/resume test

Captures pre-suspend state, suspends via RTC alarm, then checks GPU communication on resume.

sudo ./test_suspend.sh        # suspend for 10s (default)
sudo ./test_suspend.sh 30     # suspend for 30s

watchdog.sh — continuous GPU monitor

Polls nvidia-smi every 30s. On failure, automatically captures diagnostics.

sudo ./watchdog.sh

Or install as a systemd service:

sudo cp gpu-watchdog.service /etc/systemd/system/
sudo systemctl daemon-reload
sudo systemctl enable --now gpu-watchdog.service

Logs are saved to watchdog_logs/.

Tested On

• Dell Inspiron 16 — NVIDIA GeForce RTX 4060 Max-Q + Intel Arc (hybrid)
• Ubuntu 24.04.1 LTS, kernel 6.14.0-37-generic
• NVIDIA driver 570.211.01, CUDA 12.8

References

• NVIDIA Driver Documentation — Power Management
• Arch Wiki — NVIDIA/Tips and tricks #Preserve video memory after suspend

License

MIT