How it works
The seller's test takes about 10 minutes. It puts the card under heavy, standardised load while reading its sensors every second, then turns the results into pass/warn/fail checks and an A–F grade.
What the test checks
- Idle baseline. If the card is still warm (from gaming or an earlier test), the app first waits for it to cool, up to 3 minutes. If it levels off well above a true idle, the temperature rise and clock drop are marked not reliable instead of being reported.
- Video memory test. Fills most of the free video memory with known patterns and reads them back. Any changed value is a memory error, the most common damage on worn cards.
- Compute stress test. Runs a large calculation at full load over and over and checks every answer matches the first one. Wrong answers or crashes mean an unstable card.
- Cool-down. Records how the card recovers.
Throughout, the app records temperature, core and memory clocks, fan speed, power, PCIe link and thermal throttling. These go into the temperature and clock charts on the certificate page.
Factory settings
Before testing, the app reads the card's power limit, clock offsets and locked clocks. A lowered power limit makes a worn card run cooler, and on a real card we tested it only cost about 11% performance, which a performance check alone wouldn't reliably catch. So the app checks the settings directly. If they aren't at factory defaults, it offers to reset them for the test and puts yours back afterwards. Certificates are refused for cards that aren't at factory settings.
Only the full test is certified
A quick 2.5-minute screening mode exists, but it isn't certified: on the first real card we tested, a short load showed 59 °C while the full test reached 85 °C. Some problems only show once a card is fully heated.
How the grade works
| Result | Effect on the grade |
|---|---|
| Any video memory error, wrong compute result or crash | F |
| Temperature 90 °C or more, throttling for 30%+ of the load, or fans at 0% while hot | D |
| Each warning: temperature 84–89 °C, or 78 °C+ with fans at 85%+ (little cooling headroom); throttling 10–30%; unsteady clocks or a 10%+ clock drop as the card heats; PCIe link below full width; performance below 85% of typical for the model | A → B → C → D |
"Throttling" here means the card slowing itself down because of heat. A card that limits its power at full load is doing something normal: that is shown on the certificate as an information line ("Power limiting") and isn't part of the grade. A card that is slow because of a power problem shows up in the performance check instead.
Room temperature and case airflow change temperatures, so compare cards tested in similar conditions. The per-model performance figures are preliminary and will improve as more cards are tested.
Why you can check a certificate
- Fresh one-time code. Before the test starts, the server issues a code that can be used once and expires. The finished report must contain it, so an old report can't be submitted again.
- Fingerprint. The report carries a SHA-256 fingerprint of its contents. Any edit after the test, even one digit, makes it fail.
- Signature. The server checks the report, refuses quick tests, simulated cards and non-factory settings, and signs the certificate with its private Ed25519 key. That key never leaves the server. Anyone can check the signature against the public key below.
- A person behind it. Each certificate belongs to a seller account with a name and join date, so it traces back to someone, and accounts that abuse the service can be blocked.
- The buyer check. The card's unique ID, board serial (if reported), model, board maker and memory size are part of the signed certificate. The buyer's app compares them with the card in their PC.
All of this shows a report wasn't changed after the test and that the card matches. It can't show that the seller's PC ran the app unmodified. See the limits.
This server's public key
Key ID fba1572a05a995da
qwdJWIlGH0DM0h45bc1fc3K4bKkuOE/uazxSYI7pi0w=
Also available as JSON at /pubkey.