NVIDIA Xid 79: GPU Has Fallen Off the Bus
Preserve evidence, isolate the affected GPU and investigate PCIe accessibility without assuming a single root cause.
On this page
Problem
A GPU becomes inaccessible and the driver reports Xid 79. Jobs may fail and the device may no longer be usable through the normal management interface.
Conclusion
NVIDIA describes Xid 79 as the driver finding the GPU inaccessible over PCI Express. It is a symptom, not proof of one failed component. Drain affected work and capture diagnostics before recovery removes useful evidence.
Environment
Record GPU identifiers, PCI bus addresses, driver and kernel versions, motherboard or server model, power configuration, recent changes and incident time. No specific hardware incident has been reproduced for this article.
Symptoms
Look for the Xid code and its timestamp in kernel logs. Correlate it with scheduler failures and device visibility. We do not provide fabricated example logs.
Cause
The error identifies lost access to the GPU. Establishing why requires platform evidence; a model parameter change alone is not an adequate root-cause explanation.
Solution
Collect read-only diagnostics while the system remains available.
nvidia-smi -q
journalctl -k -b --no-pager
lspci -nn
Preserve the incident window, including preceding PCIe or platform errors. Stop scheduling new work on the affected node. Follow the server vendor's maintenance procedure for physical inspection or power-related checks. Arrange recovery in a maintenance window and involve support when the problem recurs.
Verification
This article provides a verification procedure; no hardware measurements are claimed.
For this fault, recovery evidence matters more than throughput: confirm device enumeration, management access and a controlled workload after recovery. Monitor for recurrence and retain the original incident bundle.
Caveats
A successful reboot does not establish the root cause. Do not reseat or service energized hardware. Review diagnostic bundles for sensitive paths and identifiers before sharing them.
Build an incident timeline
Write down the first failure time, the affected PCI bus address and the workload that was running. Then compare the application log, kernel log and scheduler events around that same interval. A final application exception may occur after the device has already disappeared, so it should not automatically be treated as the initiating cause.
| Evidence | Why retain it | What it cannot prove alone |
|---|---|---|
| Kernel Xid entry | Identifies the reported device-access fault | Which physical component failed |
| Earlier PCIe events | May locate a link problem before the Xid | That replacing the GPU will fix it |
| GPU UUID and PCI address | Keeps observations attached to the same device | That enumeration is stable after reboot |
| Recent configuration changes | Narrows a controlled comparison | That the most recent change caused the fault |
The same numerical GPU index can refer to a different device after enumeration changes. Prefer a UUID or bus address when correlating evidence. If a diagnostic command stalls, record that fact and continue collecting available host logs rather than repeatedly submitting commands to an unresponsive device.
Define recovery before returning traffic
After recovery, verify device enumeration, management access and a small controlled workload before restoring ordinary scheduling. Record the duration and workload used for this check. A successful reboot establishes restored access at that moment; it does not establish a root cause or guarantee that the incident cannot recur.
Keep two notes in the incident report: what restored service and what evidence explains the failure. If the cause remains unknown, say so. Repeated faults should be escalated with the incident bundle and hardware identifiers through the vendor's support process. Never inspect connectors or reseat a card while equipment is energized.
References
Revision note
September 14, 2026: added a worked decision framework and explicit verification limits. Original publication date retained.