HS-BM / RTX 5090
The single-RTX-5090 burn-in, published.
Two full days. One workload the rig was sold for — vLLM serving a 7B FP16 model on a fixed batch and context length. Wall-meter watts, GPU junction temperature, fan curve and tok/s, captured at one-second cadence by an instrumented rig, then reduced to the numbers a buyer can quote.
What you can quote
The headline tok/s on this page is the lowest per-hour mean sustained across the 48-hour window. That is the number a 72-hour research job will actually see — not a marketing ceiling, not a 5-minute peak.
Junction temperature is the live 1 Hz telemetry. Wall power is read off the AC mains. There is no smoothing, no cherry-picked steady-state window, and the rig that produced these numbers is owned by Helmstone.
HS-BM-02 / Headline
The numbers a buyer can quote.
The full table of measurements from the published 48-hour run. Peak and mean values derived from the 1 Hz raw ledger; sustained and aggregate values sourced from the rig-reported headline stats. Public chart (below) is downsampled to ≤8 000 points.
HS-BM-03 / 48-hour trace
Sustained tok/s, junction °C, wall W.
Three signals over the full window. The chart shows the published run; raw per-second ledger lives on the private archive and is downsampled to keep public load low.
Acoustic envelope
32 dBA · 41 dBA.
Idle (ambient) and sustained-load SPL at 1 m, measured against a Class-1 reference. Office-floor friendly, not data-center loud.
Fan envelope
PWM · broad plateau.
Fan PWM plateaus once the loop is warm; the numbers above are the acoustic load that plateau produces, not a worst-event spike.
Read the fan rows
In the table above.
Fan % at idle (first 60 s) and the sustained median over the 48-hour window sit in the headline table. PWM is a proxy for acoustic load, not a measured SPL.
HS-BM-04 / Workload & spec
What the rig was running.
The model, schema and capture window behind the published numbers. A reproducible spec table so the workload can be re-run on equivalent hardware.
HS-BM-05 / Methodology
How the headline number is chosen.
The published tok/s is a worst-hour figure. We do not smooth the curve, we do not cherry-pick the steady-state window, and we own the rig the data comes from.
Pre-sales pins the SKU
A pre-sales engineer locks the target workload before the rig is built. The 7B inference rig validates against 7B inference; the 70B rig, against 70B.
48-hour loop
Workload runs continuously at a fixed batch and context length. The rig polls thermocouples, wall-meter, nvidia-smi and the LLM /metrics endpoint each second.
Worst-hour publish
The published tok/s is the lowest per-hour mean across the window. The published junction temp is the 30-minute mean at peak load, not the absolute max.
Canonical sign-off
An engineer promotes the captured run to the singleton pointer at /api/benchmarks/canonical. That pointer is what sales attaches to every tier-one quote.