Diagnose 'Xid 79: GPU has fallen off the bus' crashes on NVIDIA
Read both BBS threads, NVIDIA open-gpu-kernel-modules#1151 and its comment 5502412431, and the Arch wiki NVIDIA/Troubleshooting raw text. Three defects. (1) The solved GTX 970 thread does not support 'disable ASPM, which helped': the poster lists 'Disable ASPM' among fixes already tried, and the kernel line in the first post already has pcie_aspm=off and NVreg_EnableGpuFirmware=0 while it was still crashing. The final post changed cleaning, cabling, a rebuilt boot partition and kernel reinstall and says 'Im not sure what fixed it'. A GTX 970 is Maxwell, which has no GSP, so that parameter did nothing there. (2) The fix omits the workaround the wiki's own 'Xid 79' section gives from #1151: nvidia-smi --lock-memory-clocks with a min,max range, and the laptop thread's poster also reports that locking minimum clocks to keep the GPU out of P5 works, so 'remains unresolved' is incomplete. (3) The GSP step can never apply on Omarchy 4: /usr/share/omarchy/install/hardware/nvidia.sh installs nvidia-open-dkms for device IDs >= 0x1e00 (Turing+, GSP required) and nvidia-580xx-dkms only for Maxwell to Volta, which have no GSP. nvidia-utils ships nvidia-smi, nvidia-bug-report.sh and nvidia-persistenced.service (pacman -Ql). Nothing was exercised.
A locked memory or graphics clock floor raises idle power draw and heat, more so the higher the floor, and noticeably on laptops. pcie_aspm=off also raises idle power use. Opening a desktop case to re-seat power cables should be done with the machine unplugged.
Symptom
The screen freezes or goes black, the session dies, or the dGPU vanishes, sometimes under load and sometimes seconds after boot or at idle. journalctl -k -b -1 shows:
NVRM: Xid (PCI:0000:27:00): 79, GPU has fallen off the bus.
sometimes followed by Xid 154: GPU Reset Required. nvidia-smi then fails until a reboot.
Cause
Xid 79 means the driver tried to reach the GPU over PCI Express and found it inaccessible. NVIDIA's description and the Arch wiki list hardware causes (power supply, cables, the PCIe connection) and driver issues, so it is a symptom with several causes rather than one bug. In the solved Arch forum thread (GTX 970), the crashes continued with pcie_aspm=off and nvidia.NVreg_EnableGpuFirmware=0 already set, and stopped after the case was cleaned, the GPU power cabling was rewired and the boot setup was rebuilt. The poster could not say which change fixed it. A GTX 970 has no GSP, so the firmware parameter did nothing there. In a second thread, an Ampere laptop GPU dropped off the bus every time it moved down from the P5 power state, pcie_aspm=off and both NVreg_DynamicPowerManagement values did not help, and locking the minimum clocks so the GPU never entered P5 worked around it. NVIDIA open-gpu-kernel-modules#1151 reports Xid 79 on desktop cards with no warning at any load, and a comment there reports that locking the memory clock range with nvidia-smi -lmc stopped it. The Arch wiki repeats that workaround.
Fix
**1. Rule out hardware first.** Check temperature and power under load, clean the cooler, and on a desktop re-seat the GPU and its power connectors:
nvidia-smi --query-gpu=temperature.gpu,power.draw,pstate --format=csv -l 2
journalctl -k -b -1 | grep -E 'Xid|NVRM'
**2. Lock the memory clock range** (the Arch wiki workaround, from NVIDIA open-gpu-kernel-modules#1151). List the supported memory clocks, then allow the lowest up to one step below the highest:
nvidia-smi -q -d SUPPORTED_CLOCKS | grep Memory
sudo nvidia-smi --lock-memory-clocks 405,9251 # use your own lowest and second-highest values
This resets at reboot. Undo it with sudo nvidia-smi --reset-memory-clocks. On Hopper data-centre cards use --lock-memory-clocks-deferred instead. On a laptop whose GPU dies when leaving a specific P-state, the same thread reported that a higher minimum graphics clock also helps:
nvidia-smi -q -d SUPPORTED_CLOCKS | grep -E 'Graphics|Memory'
sudo nvidia-smi --lock-gpu-clocks 800,2100 # example values from the wiki, use yours
To keep the lock across reboots, the wiki uses a oneshot unit. This is the same on Omarchy 4 and plain Arch:
sudo tee /etc/systemd/system/nvidia-clocks.service >/dev/null <<'EOF'
[Unit]
Description=Lock NVIDIA memory clock range
Requires=nvidia-persistenced.service
After=nvidia-persistenced.service
[Service]
Type=oneshot
ExecStart=/usr/bin/nvidia-smi --lock-memory-clocks 405,9251
RemainAfterExit=yes
[Install]
WantedBy=multi-user.target
EOF
sudo systemctl enable --now nvidia-persistenced.service nvidia-clocks.service
**3. Optionally try disabling PCIe ASPM.** It did not stop the crashes in either Arch forum thread, so try it only after the steps above.
Omarchy 4, in your own drop-in, with +=:
sudo tee -a /etc/limine-entry-tool.d/zz-local.conf >/dev/null <<'EOF'
KERNEL_CMDLINE[default]+=" pcie_aspm=off"
EOF
sudo limine-mkinitcpio && sudo limine-update
sudo reboot
Plain Arch, GRUB: add pcie_aspm=off to GRUB_CMDLINE_LINUX_DEFAULT in /etc/default/grub, then sudo grub-mkconfig -o /boot/grub/grub.cfg.
**4. GSP firmware (nvidia.NVreg_EnableGpuFirmware=0) only matters with the closed kernel modules on a Turing or newer card.** It does not apply on Omarchy 4: Turing and newer cards get nvidia-open-dkms, whose modules require GSP, and Maxwell to Volta cards get nvidia-580xx-dkms, and those GPUs have no GSP. On plain Arch with the closed modules on a Turing or newer card, add it the same way as step 3. See nvidia-gsp-firmware-crashes.
**5. Collect a report** if it persists, for NVIDIA's forum or the open-gpu-kernel-modules tracker:
sudo nvidia-bug-report.sh
Verify
Several days of normal use and a sustained load test pass with no Xid lines in journalctl -k. nvidia-smi -q -d CLOCK shows the locked range is active after a reboot if you added the unit, and cat /proc/cmdline contains any kernel parameters you added.
Sources
- https://bbs.archlinux.org/viewtopic.php?id=304020
- https://bbs.archlinux.org/viewtopic.php?id=313284
- https://wiki.archlinux.org/title/NVIDIA/Troubleshooting
- https://github.com/NVIDIA/open-gpu-kernel-modules/issues/1151
- https://github.com/NVIDIA/open-gpu-kernel-modules/issues/1151#issuecomment-5502412431
- https://github.com/omacom/omarchy/blob/quattro/install/hardware/nvidia.sh