eng-0148
window reconstructed
citable URL: https://halobench.com/records/eng-0148/ — this address never moves; the anchor /records/#eng-0148 keeps resolving
159.3 Whmean 176.77 W − 10.1 W idle floor (box-idle-all-empty) = 166.67 W attributable
window 2026-08-17T08:37:58Z → 2026-08-17T09:32:02Z · 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, qwen38-27b Q8_0 ROCm d262144 model-max (cfg-0103), covering both the prefill row (run-0375) and decode row (run-0376) from the single llama-bench invocation. Session window reconstructed from ~/bench-results/depth-backfill-20260817/qwen38-27b/d262144.started (08:37:58Z) and the output JSON's file mtime (09:32:02Z); same reconstructed-not-recorded provenance as eng-0147. Counter interpolated from raw ~10s-resolution HA history samples: start 24.955448 kWh (08:37:58Z), end 25.114737 kWh (09:32:02Z). The deepest cell in this backfill (~54 min wall time), consistent with the run-0375 comment's own wall-time estimate.
Wh, never joules — read as a cumulative-counter difference across the window; delta over the idle baseline is the only figure that means anything