eng-0116
window recorded
citable URL: https://halobench.com/records/eng-0116/ — this address never moves; the anchor /records/#eng-0116 keeps resolving
648.6 Whmean 146.23 W − 10.1 W idle floor (box-idle-all-empty) = 136.13 W attributable
window 2026-08-16T02:32:56Z → 2026-08-16T06:59:04Z · provenance recorded
method wall-meter · aihydra energy sensor via Home Assistant (counter-difference on sensor.hardware_ai_hydra_energy, raw ~10s history samples interpolated to the run-meta window) · no contention
per task 32.43 Wh · 0.9826 p at 30.3 p/kWh
covers run-0316
aihydra run-meta.jsonl entry nemotron3-super-fullbench/tau2-full-26task (rc=0 — all 26 requested tasks completed cleanly, not a wall-bound cut). Joined same-session: counter interpolated from raw ~10s-resolution HA history samples at each boundary — start 20.99128201 kWh (03:32:56+01:00 local / 02:32:56Z), end 21.63988203 kWh (07:59:04+01:00 local / 06:59:04Z). wh_per_task is Wh-per-CORRECT-ANSWER: wh_total / tasks_passed (20 of 26, run-0316.metrics). WHOLE-SESSION energy, not agent-only — the window includes simulator (user_llm) wait time end to end, directly comparable on that denominator to qwen38-27b's eng-0081 and deepseek-v4-flash's eng-0107. Mean power (146.23 W) sits between the throughput matrix's rocm-d32768 cell (147.74 W, eng-0113) and vulkan-d32768 cell (149.47 W, eng-0115) — consistent with a mostly-decode-bound agentic workload at the same serving depth band.
Wh, never joules — read as a cumulative-counter difference across the window; delta over the idle baseline is the only figure that means anything
Cited by — computed at build time, never stored
model pages nemotron3-super