eng-0150
window reconstructed
citable URL: https://halobench.com/records/eng-0150/ — this address never moves; the anchor /records/#eng-0150 keeps resolving
55.41 Whmean 179.37 W − 10.1 W idle floor (box-idle-all-empty) = 169.27 W attributable
window 2026-08-17T09:48:14Z → 2026-08-17T10:06:46Z · provenance reconstructed
method wall-meter · aihydra energy sensor via Home Assistant (counter-difference on sensor.hardware_ai_hydra_energy, raw ~10s-resolution history samples, linearly interpolated to the session window) · no contention
Deadline-critical batch join (protocol §11) for the 2026-08-17 depth-ladder backfill, ornith-35b Q8_0 ROCm d262144 model-max (cfg-0104), covering both the prefill row (run-0379) and decode row (run-0380) from the single llama-bench invocation. Session window taken from the launch timestamp embedded in the run-0380 record's own comment (09:48:14Z launch) cross-checked against the ~/bench-results/depth-backfill-20260817/ornith-35b/d262144.json file mtime (10:06:46Z, used as the end boundary); reconstructed, not runmeta-recorded (same gap as eng-0147..eng-0149). Counter interpolated from raw ~10s-resolution HA history samples: start 25.152866 kWh (09:48:14Z), end 25.208271 kWh (10:06:46Z). Near-identical mean power to eng-0149 (179.37 W vs 179.34 W) -- same model/backend/build, only depth differs, consistent with a compute-bound GPU draw that doesn't move much with depth at this model's cheap-KV footprint.
Wh, never joules — read as a cumulative-counter difference across the window; delta over the idle baseline is the only figure that means anything