gpucheck stress-tests an NVIDIA GPU for ten minutes and grades it against healthy cards of the same model — one command, free, open source. Selling the card? Get the result as a $29 signed certificate.
Apache-2.0 · Linux + NVIDIA · no account needed
How it works
No benchmarks to interpret, no spreadsheets of clock speeds. gpucheck measures how the card behaves under sustained load and tells you in plain language.
gpucheck records an idle baseline, then runs a sustained matrix-multiply load for about ten minutes while sampling temperature, clocks, power, and throttle reasons.
If you choose to upload, only result.json is sent — the telemetry and a
salted-hash card identifier. You get a public result page and a delete token.
Sustained TFLOPS is compared against a reference set of healthy cards of the same model, and problems are explained: thermal throttling, power capping, memory errors, below-spec throughput.
What it sees
A card can look fine for thirty seconds and fall apart at minute four. Here's a real failure pattern gpucheck catches: temperature climbs, the card starts thermal-throttling, clocks drop — and so does throughput.
swipe → to follow the run
The certificate — $29
Run gpucheck --certify and the run is bound to a fresh
one-time challenge from our server, so the result could not have been precomputed or
replayed. The graded result is signed with Ed25519 and anyone can verify it online.
A certificate raises the cost of lying about a GPU. It is not proof. We'd rather tell you that plainly than sell you certainty that doesn't exist.
Sample for illustration — layout and fields match a real certificate.
Pricing
Checking your own hardware costs nothing, ever. Pay only when you need a result someone else can trust.
Payments processed by Paddle.com as merchant of record. VAT/sales tax handled at checkout where applicable.
Install
Linux with an NVIDIA GPU and driver installed. Nothing is uploaded unless you say so.
# quick check with uv (no install)
uvx --from "gpucheck[load]" gpucheck --quick
# or install with pip
pip install "gpucheck[load]"
# certified run ($29, pays at upload)
gpucheck --certify
FAQ
Linux with NVIDIA GPUs only, for now. gpucheck reads telemetry through NVML, so it needs the NVIDIA driver installed. No Windows, macOS, AMD, or Intel support yet.
Partially. On MIG-partitioned GPUs gpucheck can take measurements, but it won't grade them — a MIG slice isn't comparable to the healthy-card reference set for the full model.
Uploading is opt-in, and only result.json is sent: the telemetry
samples, GPU model, driver version, and findings. The card identifier is a salted hash —
the raw GPU UUID never leaves your machine. No hostname, no user name, no other system
information.
Yes. Every upload returns a delete token; use it to remove the result page at any time. Keep it safe — there are no accounts, so the token is the only way to prove the result is yours.
No, and we won't pretend otherwise. It proves a fresh, untampered, server-graded run happened on a card with that identifier. It cannot stop someone who controls the machine from fabricating telemetry, or from re-running until they like the result. It raises the cost of lying; it is not proof.
Yes — the CLI is open source under Apache-2.0, and health checks, grades, and result pages cost nothing. The only paid thing is the $29 signed certificate.