eng-0122
window recorded
citable URL: https://halobench.com/records/eng-0122/ — this address never moves; the anchor /records/#eng-0122 keeps resolving
48.57 Whmean 91.25 W − 10.1 W idle floor (box-idle-all-empty) = 81.15 W attributable
window 2026-08-16T08:21:07Z → 2026-08-16T08:53:03Z · provenance recorded
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 run-meta window) · no contention
covers run-0327
aihydra laguna-s21-fullbench throughput matrix cell vulkan-d131072 -- CELL FAILED, both allowed attempts lost the GPU device (rc=134, vk::DeviceLostError; run-0327). This is WASTED energy, not a throughput result: 48.57 Wh spent across two attempts (1035s + 881s) that each aborted before producing a decode/prefill number. Counter interpolated from raw ~10s-resolution HA history samples pulled fresh this session: start 21.75803541361878 kWh, end 21.80660287926063 kWh. Recomputed independently from HA raw history; matches the MODEL-PAGE.md draft to within rounding (48.57 Wh / 91.3 W). Recorded per protocol's "publish the frontier, not the winner" -- a failed cell carries its own energy cost and that cost is part of the record.
Wh, never joules — read as a cumulative-counter difference across the window; delta over the idle baseline is the only figure that means anything