Saltar al contenido
MZGrid

Esta página solo está disponible en inglés. La versión en inglés es la que prevalece.

Benchmark & verification

How MZGrid verifies provider hardware and work: registration benchmarks, usage reports and canary spot checks.

MZGrid verifies that providers have the hardware they claim and do the work they are paid for. Verification has four layers; failing the later ones leads to slashing.

1. Registration benchmark

When you run mzgrid-provider register (or benchmark), the agent measures FP16 dense matrix-multiply throughput on every GPU and submits the results to POST /v1/providers/benchmark, signed by your provider key:

JSON
{
  "machineId": "m-fra-01",
  "results": [
    { "gpuUuid": "GPU-4f2b9c1e-7a0d-4e5b-9c3a-2d8e1f6a7b90", "tflopsFp16": 158.4, "memBandwidthGBs": 978.2, "durationMs": 41250 }
  ],
  "signature": "0x9a1f...c41b"
}

The measured TFLOPS must be between 70% and 115% of the catalogue reference for the claimed SKU. Outside that window the machine is marked unverified and does not receive jobs.

SKUReference FP16 TFLOPSAccepted window
b20022501575 – 2587.5
h200989692.3 – 1137.35
h100-sxm989692.3 – 1137.35
h100-pcie756529.2 – 869.4
a100-80312218.4 – 358.8
a100-40312218.4 – 358.8
l40s362253.4 – 416.3
rtx6000-ada364254.8 – 418.6
rtx5090210147 – 241.5
a6000155108.5 – 178.25
rtx4090165115.5 – 189.75
rtx4080s10472.8 – 119.6
rtx30907149.7 – 81.65

The lower bound tolerates power limits, thermals and driver differences; the upper bound catches mislabelled hardware and implausible results — for example a faster GPU registered under a cheaper SKU, or a reported number no card of that SKU can reach. Always register each GPU under the SKU that detect reports.

Getting a good result

  • Run the benchmark on an idle GPU, with no other processes using it.
  • Make sure the card is not thermally throttling: check nvidia-smi -q -d PERFORMANCE.
  • Do not set aggressive power limits below the default.
  • Use a current driver supported by your CUDA runtime.

Benchmark results also feed the benchmark consistency component of your reputation: repeated runs that agree with each other and with the reference score higher.

2. Usage reports

While a job runs, the agent sends POST /v1/providers/jobs/:id/usage every 60 seconds with GPU seconds and utilisation samples. The coordinator requires samples at least every 90 seconds. Short gaps are tolerated; a gap of more than 10 minutes while a job is running is treated as the machine being offline: the job is failed with a full refund to the customer and the provider is slashed for OfflineDuringJob.

The agent also sends a heartbeat every 30 seconds with per-GPU utilisation, memory and temperature, which feeds your uptime score.

3. Spot checks

2% of jobs are canary jobs: the payload has a known expected output hash. When the job finishes, the agent reports a resultHash. A mismatch means the work was not performed as specified and is slashed as VerificationFailed — 100% of your locked balance. Canary jobs look and are paid like any other job, so the only way to pass them is to run every job honestly.

4. Deadlines

An accepted job that is not finished by its on-chain deadline is treated as abandoned and slashed. Keep the agent running, keep the machine online and do not take machines down for maintenance while they have jobs. To take a machine out of service, stop the agent from accepting new work and wait for running jobs to end.