eng-0149
window reconstructed
citable URL: https://halobench.com/records/eng-0149/ — this address never moves; the anchor /records/#eng-0149 keeps resolving
37.11 Whmean 179.34 W − 10.1 W idle floor (box-idle-all-empty) = 169.24 W attributable
window 2026-08-17T09:35:22Z → 2026-08-17T09:47:47Z · 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 d204800 (cfg-0104), covering both the prefill row (run-0377) and decode row (run-0378) from the single llama-bench invocation. Session window taken from the launch/landing timestamps embedded in the run-0378 record's own comment (09:35:22Z launch, 09:47:35Z decode landing) cross-checked against the ~/bench-results/depth-backfill-20260817/ornith-35b/d204800.json file mtime (09:47:47Z, used as the end boundary); reconstructed, not runmeta-recorded (same gap as eng-0147/ eng-0148). Counter interpolated from raw ~10s-resolution HA history samples: start 25.115512 kWh (09:35:22Z), end 25.152626 kWh (09:47:47Z). Cheapest cell of the backfill by wall time (~12 min) -- sparse A3B-active MoE, per the run-0378 comment.
Wh, never joules — read as a cumulative-counter difference across the window; delta over the idle baseline is the only figure that means anything