False hardware attribution
Prevent application, driver, power, thermal, riser and fabric faults from being presented as a defective board.
For hardware support and integration teams
OEM, systems integrator and GPU RMA workflows
A useful GPU RMA workflow must answer two questions before generating paperwork: does the evidence meet a hardware threshold, and does the fault follow the card rather than the slot, node, power path, driver or workload?
Denpex organizes Xid, ECC, NVLink, device identity, recurrence and verification evidence into a vendor-facing incident path. Ambiguous cases remain field-diagnosis candidates, while software-owned failures are explicitly kept out of the RMA queue.
Who this is for
- GPU OEM support teams
- Systems integrators
- Hardware qualification engineers
- Field service engineers
- Data center operations teams
- GPU cloud escalation teams
The failure surface
These incidents look adjacent in the final alert, but they require different owners and different proof before recovery.
Missing device identity
Attach GPU UUID, serial, PCI bus ID, host and firmware context before the hardware is removed.
Evidence lost during reset
Preserve kernel, BMC, Xid, ECC and recurrence evidence before power cycling clears the most useful state.
Vendor support loops
Package completed checks and the remaining uncertainty so the next escalation does not repeat basic collection.
The operating loop
- 1
Classify the error
Map the initiating signature to application, environment, fabric, configuration, hardware or unresolved ownership.
- 2
Collect board and host evidence
Record device identity, health state, recurrence and shared-node signals before hardware movement.
- 3
Run the ownership control
Test whether the failure follows the card, slot, node, workload or software image.
- 4
Generate the support packet
Include the verdict, evidence checklist, commands, timestamps and unresolved items in the vendor case.
What Denpex contributes
- NVIDIA Xid classification with explicit hardware and non-hardware fault classes
- Evidence-gated RMA payloads using node and GPU incident history
- Verification commands and missing-evidence instructions before escalation
- Fleet metrics for repeated hardware events and measured downtime
Measure this in an evaluation
Use your incidents and timestamps. These are measurement definitions, not Denpex customer-outcome claims.
| Metric | Measurement |
|---|---|
| False RMA avoidance | Suspected boards cleared by a controlled test that identified slot, node, power, thermal, driver or workload ownership. |
| Vendor acceptance cycle | Support exchanges required before the case contains accepted device identity, health evidence and reproduction data. |
| Time to hardware verdict | Elapsed time from first incident to RMA candidate, field diagnosis or not-RMA-worthy classification. |
| Repeat device incidents | Verified hardware-family events per GPU UUID after repair, replacement or return to service. |
Start with the technical evidence
NVIDIA Xid decoder
See fault class, first action and RMA relevance for common driver-reported codes.
GPU reliability evidence
Review sourced failure surfaces across accelerators, memory and interconnects.
Xid 79 investigation
Separate a missing GPU from power, thermal, PCIe, slot and board ownership.
Reliability and RMA metrics
Export per-node events, MTTR and measured cost data to existing reporting.
Frequently asked questions
Does Xid 79 automatically justify a GPU RMA?
No. It proves the driver lost communication with the device, but a power event, thermal cutoff, riser, slot, baseboard or dead card can all produce that symptom. The cheapest discriminator is whether the failure follows the card into a known-good slot.
What evidence should be collected before a GPU reset?
Preserve the kernel log, NVIDIA bug report, GPU UUID and serial, PCI bus ID, ECC and row-remapping data, NVLink state, BMC power and thermal events, node topology, workload timestamp and any simultaneous failures on peer GPUs.
Can Denpex generate a vendor support case?
The fleet workflow can assemble an evidence payload and support-ready text from the available incident data. Missing fields remain marked with collection instructions, and the verdict can explicitly say field diagnosis first or not RMA-worthy.
How do we measure whether the workflow saves money?
Track actual boards withheld after a non-card cause was verified, support exchanges per accepted case, hardware verdict time and downtime by GPU UUID. Keep measured outcomes separate from modeled replacement or GPU-hour costs.
Prove it on incidents your team already resolved
Freeze the expected owner and outcome, replay the evidence, then compare operator time and decision quality with your current process.
Build the evaluation plan