You rent a "RTX 4090, 24 GB, 128 GB RAM" by the hour. The machine boots, nvidia-smi prints a GPU name, and you start a training run. But is it really the card in the listing, with the memory, the CPU, the RAM and the disk you are paying for?
On GPU marketplaces the answer is usually yes. Sometimes it isn't: a host lists the wrong model, a node shows up with one GPU fewer than advertised, VRAM is smaller than the listing, or the disk is half the size. You find out hours later, when the job is slow or runs out of memory.
So we wrote a short, read-only check you run yourself, in one command, on any Linux machine with an NVIDIA GPU. It is not tied to us: it works the same on any provider.
The one-liner
curl -fsSL https://voltagegpu.com/hwcheck.sh | bash -s -- --gpu "RTX 4090" --gpus 1 --vram 24 --ram 128 --disk 500
You pass what the listing says. The script measures the machine and prints one line per item, then a verdict:
RESULT: MATCHES THE LISTING
or DOES NOT MATCH, with the line that failed. The exit code is 0 when everything matches and 2 when something doesn't, so you can put it at the top of a startup script and stop paying the moment a machine is wrong.
What it checks
- GPU model. Every token of the listed name that carries a digit (
4090,h100,a6000,80gb) must appear in the name the driver reports. A card sold as a 4090 that reports a 3060 Ti fails here. Mixed GPUs in one machine are flagged. - GPU count. The number of GPUs the driver sees, against the listing.
- VRAM per GPU. The smallest card's memory, against the listing, with a small tolerance.
- Real CUDA work. If the image ships PyTorch, it runs a matrix multiplication on every GPU, so a card that is listed but unusable fails.
- CPU, RAM and disk. Measured against what you passed.
- Network. One 25 MB download from Cloudflare to measure speed. Skip it with
--no-net.
Nothing about the machine is sent anywhere. The script is 149 lines of bash you can read before you run it: https://voltagegpu.com/hwcheck.sh
What it does not prove
This catches mislabeled machines, wrong counts and undersized nodes, which is what goes wrong in practice. It does not defend against a host determined to lie: someone who controls the machine can, in principle, make the driver report a different name.
If you need proof against the operator itself, you need hardware attestation: a confidential VM where the CPU signs a quote of the trust domain and NVIDIA's service signs the GPU's evidence, both bound to a challenge you chose. We wrote about that, with an open-source verifier, here: https://voltagegpu.com/blog/two-proofs/voltage-verify
How we use it
On VoltageGPU, the standard GPU pages give you this command for the machine you pick. If a pod doesn't match its listing, send the output to support and we look at it the same day. Standard GPUs carry no enclave and no attestation, so they are for data that isn't sensitive; for anything else, use a confidential VM.
Live standard GPU stock: https://voltagegpu.com/standard-gpus