Certificate
The Known Good certificate
One certificate per unit, tied to its serial. It records what was tested, with which tool, the numbers produced, and who signed it. The same protocol applies to every unit, and the full commands are published in the test guide.
What is tested
Identity and labels
Model, part number and serial are read from the board and from nvidia-smi and must match each other and the labels. Photos of the labels go on the certificate.
Visual inspection
Connectors, heatsink, fins, screws and the PCB edge are inspected and photographed. Damage, missing parts and signs of rework are recorded.
Firmware and VBIOS
VBIOS version, InfoROM and ECC state are recorded with nvidia-smi -q. Nothing is flashed without saying so on the certificate.
DCGM diagnostics
NVIDIA DCGM diagnostics at level 3 (dcgmi diag -r 3), covering PCIe, memory, the SM stress test and targeted power. Every plugin must pass.
Memory test
Full memory test across the HBM or GDDR, with ECC counters read before and after. Retired pages and remapped rows are recorded.
NVLink and NVSwitch bandwidth
Where the unit has NVLink, nccl-tests all_reduce_perf measures bus bandwidth across all GPUs. The figure is recorded against the expected value for the platform.
Soak run
Sustained load for the stated hours with temperature, clocks and power logged once a second. The record shows any thermal or power throttling.
Power draw
Board power under load against the rated TDP for the configuration.
Performance baseline
A GEMM run (and HPL for systems) compared with the datasheet figure, so a slow unit shows up as slow.
OEM warranty lookup
The serial is checked against the manufacturer's warranty records where a lookup exists. The result and the date of the check are recorded.
Provenance
Where the unit came from, in general terms, and whether it was in a production cluster. Supplier and client names are never published.
Who tests
Units are tested by me, Christian Busch, unless the certificate names another tester. The tester is named on every certificate.
Who signs
I sign every certificate. My signature states that I ran or supervised the tests listed, on the unit with the stated serial, on the stated dates, and that the results shown are the results recorded.
How to verify
Each certificate has an ID and a page at knowngoodcompute.com/verify/ followed by that ID. The page shows the model, part number, the last four characters of the serial, the test date, the tester, the signatory and the SHA-256 fingerprint of the PDF. If the fingerprint of your PDF does not match, it is not the certificate that was issued.
Sample certificate
The sample certificate KG-SAMPLE-0001 shows the layout with example values. It does not describe a real unit. Download the sample PDF.
Limits
A certificate records how one unit performed on the stated dates under the stated tests. It does not predict future failures, it is not a warranty, and it does not replace the Terms of Sale or any manufacturer warranty. Tests that could not be run on a given unit are marked as not run, never left out.