Identification
Manufacturer, model, GPU type, and GPU count, taken from the hardware. Serial numbers are copied when the labels are present.
Home Quality
Quality monitoring
Checks and documentation for every offer.
A quality report is the written record of the inspection we ran on the server, node, or rack in an offer: a gpu-burn stress test, NVLink status, and ECC error counts for each GPU, plus identification and condition. It travels with the conversation before shipment.
| GPU | gpu-burn | NVLink | ECC corr. | ECC uncorr. |
|---|---|---|---|---|
| GPU 0 | PASS | all up | 0 | 0 |
| GPU 1 | PASS | all up | 0 | 0 |
| GPU 2 | PASS | all up | 0 | 0 |
| GPU 3 | PASS | all up | 0 | 0 |
| GPU 4 | PASS | all up | 0 | 0 |
| GPU 5 | PASS | all up | 2 | 0 |
| GPU 6 | PASS | all up | 0 | 0 |
| GPU 7 | PASS | all up | 0 | 0 |
The panel shows example output in the format of a report. It is not a completed inspection of a listed unit and it is not a certificate.
Checks
What the inspection covers.
The same headings are used across servers, nodes, and racks. The report states which of them were completed for that unit.
Visual condition
Chassis, connectors, fans, and labels, including obvious damage. The notes stay with the unit they describe.
Power-on
When the hardware can be powered in the inspection setup, the report records whether it powers on and whether the GPUs are recognized.
Stress test
Each GPU is put under sustained full load with gpu-burn. The report records whether every GPU completed the run and any errors it produced.
NVLink and ECC
NVLink status is read for each GPU, so a down or degraded link shows up. Corrected and uncorrected ECC error counts are written down as found.
Limits
If a label is missing, a unit cannot be powered, or a rack is offered as a lot, the report says which checks stopped there.
Sample report structure
The fields a report is built from.
This is the shape of the document. Values are filled in for the specific hardware in an offer. Nothing in this table is a listed unit.
| Field | What is recorded |
|---|---|
| Report reference | An internal reference for this inspection. |
| Date | The date the checks were written down. |
| Item | Server, node, or rack. |
| GPU model and count | The model and count identified on the hardware. |
| Serial numbers | Numbers present on labels. A missing label is noted. |
| Visual condition | Chassis, connectors, fans, and obvious damage. |
| Power-on | Whether the unit was powered, and what the check recognized. |
| Stress test | gpu-burn result for each GPU under sustained load, including errors. |
| NVLink | Link status for each GPU, including links that are down or degraded. |
| ECC errors | Corrected and uncorrected counts for each GPU, as read at inspection. |
| Photos | References to images taken for that inspection, when photos are part of the file. |
| Limits | Checks that were not completed, and the reason. |
Sample structure only. This page does not publish a completed report, a certificate, or a unit for sale.
Reading the results
What the numbers mean.
A report is only useful if you can read it. These are the three results to look at first.
gpu-burn
A pass means the GPU completed the sustained full-load run without computation errors. Any error is recorded and shown against that GPU, so one weak card in a server does not hide behind the others.
NVLink
Every expected link should show as active. A link that is down or degraded is recorded, because it limits how well several GPUs train together.
ECC errors
Corrected errors were fixed by the memory in operation and are counted. Uncorrected errors mean data was lost, and they are the figure to watch. Both counts are written down as found.
Scope of the record
What the report is for.
Buyers use the report to see what was checked. The boundaries below are part of every conversation about that document.
The record itself
The report documents identification, visual condition, power-on results when the check was run, and the limits of that inspection for the hardware in the offer.
Outside the record
The report does not issue an OEM warranty, a calibration certificate, or a statement that the hardware meets a named standard. Those documents, when they exist, come from other parties.