HS-BM / Per-tier
Real numbers per tier. Pending tiers, named.
Every Helmstone unit ships with a published benchmark. This page is the per-tier view: sustained tok/s and peak thermals for the rigs we have shipped, an honest “pending burn-in” tile for the chassis in thermal validation, and the shared protocol that ties them together.
HS-BM-MTH / Methodology
How the headline number is chosen.
The published tok/s on the cards below is a worst-hour figure, captured against the same protocol laid out here. Methodology is identical to /benchmarks/rtx-5090; tiers whose burn-in has not returned from the rig render a transparent “pending” tile rather than a fabricated number.
GPU junction temperature
The actual temperature of the silicon at the hottest spot on the die — not the case fan reading.
Four k-type thermocouples are clamped to the die hotspot, the VRAM side, the memory side, and the exhaust. The headline number is the peak across all four channels during the 48-hour window.
Sustained inference workload
A real model loop, not a synthetic stress run — so the number reflects what a 72-hour job will see.
Either vLLM serving a 7B or 70B FP16 model, or llama.cpp against the same model class. Batch and context length are pinned; throughput and tail latency are sampled against the model’s own Prometheus /metrics endpoint.
Wall-meter power
Total AC-mains draw the machine pulls from the wall — what your facility actually plans around.
Read off a Shelly EM clamp on the AC mains, or a USB-TMC power analyzer on the chassis PDU. Sampled at the same 1 Hz cadence as the GPU telemetry.
Ambient conditions
The room temperature and humidity at the bench — so the headline is reproducible in your lab, not just ours.
Logged alongside the rig the whole way through. If the bench moved on the day a buyer’s reference number was captured, that is noted on the per-unit benchmark sheet.
Acceptance pass criteria
- Pass 01No throttle eventsZero `nvidia-smi` throttle-state transitions recorded during the 48-hour window. Any single throttle event fails the unit out of the line.
- Pass 02Peak junction below the validated envelopeHeadline peak junction temperature stays under the per-SKU ceiling. VRAM and exhaust channels are checked separately to confirm the cooling profile is correct, not just the die hotspot.
- Pass 03Sustained wall-meter watts stableTotal system draw does not regress across the window. A drop is a sign of thermal compensation or throttling that defeats the burn-in; both are pass-failing.
- Pass 04Sustained tok/s within the published bandWorst-hour throughput sits inside the band published on the unit datasheet. The published number is what we ship — anything above it is engineering margin, not marketing.
- Pass 05Per-unit benchmark signedThe burn-in summary sheet is signed by an engineer before the unit is sealed. That signature is what the on-site warranty honours when a buyer opens a fault ticket.
HS-BM-GRID / Per-tier grid
One card per tier. Same protocol; honest status.
The three workstation tiers, side by side. The published tier carries the live numbers from /api/benchmark-tiers (which lifts worst-hour tok/s and peak junction temperature from the canonical singleton). Pending tiers carry chassis facts we can quote today — VRAM, tier label, GPU and workload class — and a single status word.
HS-BM-FAQ / Buyer questions
What the benchmark means for a buyer.
Use the same evidence to assess workload fit, procurement scope, support coverage, and the methodology behind each published number.