NVIDIA and AMD · one signed report per server · verify it anywhere
SiBIL measures a server, scores it from 300 to 900, and signs the result. Buyers see whether a box fits their workload before they rent it. Providers prove what they sell. Nobody can edit the number.
curl -fsSL https://sibil.dev/install.sh | sudo sh -s -- <token>A spec sheet lists the GPU. It does not list the PCIe link that trained down to x8, the 16 vCPUs that feed the card, or the host RAM that is smaller than the GPU memory. Buyers, renters and lenders find these defects after they pay.
An H200 that reaches its full reference compute, on a host that keeps it busy 66 % of the time. Every job on this server pays for the idle third.
| link | Gen5 x16 | Gen4 x8 |
| root port | 1 per GPU | 4 GPUs share one |
| power limit | 700 W | 500 W |
| host RAM | 2 TB | 512 GB, 4 of 12 ch |
Links that train down to Gen3 or x8. Eight GPUs behind one root port. A power limit below the SKU. A PCIe card sold as SXM. None of it is in the listing. All of it is in one run.
Provider-run numbers come from a run the provider chose, on a day the provider chose. A rating only counts when the seller cannot edit it and the buyer can verify it.
A small agent runs a versioned suite on the server and sends the signed results over ClutchCall MoQT. The backend verifies the proof, scores the server, and signs the report.
CPU, memory channels, every PCIe link, every GPU and NIC. A hardware ID that changes when a part changes, so a swapped card shows up.
Native tests for CPU, memory and storage. GEMM, copy bandwidth and the feed test on the GPU. The CPU re-adds GPU tile sums, so a faked run fails the proof.
Performance asks whether each part is as good as its reference. Balance asks whether the parts fit together. The weaker one sets the score.
The node signs the raw bundle. The backend verifies, scores and countersigns. A share link and a one-file verify.py check it anywhere.
| cpu | AMD EPYC 9004 · 16 vCPU | no reference yet |
| memory | 126 GiB | < GPU memory |
| gpu0 | H200 NVL · 141 GB | Gen5 x16 trained |
| power | 600 W limit | = SKU TDP |
| sig_alg | ecdsa-p256-sha256 |
| key_id | 67c7762dc90ca121 |
| payload | 8c25020a69b159ef16820a97e08139c92dec0d1… |
| valid | 30 days · spot check |
A weak host with good GPUs is a bad training server and a fine inference server. Each report scores eight workload classes from 0 to 100, so a buyer or a scheduler asks for the servers that fit the job.
Every run adds a point to the history of the server. A flag with a known fix becomes a refurbish line with the part to change, and the next run shows the gain.
Renting? See the score
before you pay.
Selling? Prove it with a badge
you cannot edit.
Lending, or buying used?
Read the signed history.
One line installs the agent. It idles until a run is due, needs no sudo after install, and reaches the backend from behind NAT. Nothing about how you run your servers has to change.
The line at the top of this page. A fleet token enrols many servers, and an image can carry the token for first boot. Root work goes through a small systemd socket with a peer check, so the agent itself never runs as root.
Results travel over ClutchCall MoQT. A lost link delays a report; it never loses one. No inbound port, no VPN.
The test suite is versioned and TUF-signed. Every report names the suite and the score model that produced it, so two servers compare only on the same version.
The GPU tools build from one source for CUDA and ROCm. The balance rules do not care who made the card.
The GPU returns tile sums the CPU checks. A nonce ties the run to the schedule. A replayed or edited bundle gets no score.
Each flag names the rule, the measured value and the threshold. See why a server lost points, and which part to fix.
Install the agent on one box. Get a signed report with a score, the flags, and the workload fit. Bring the server you are not sure about.