eng-0151
window reconstructed
citable URL: https://halobench.com/records/eng-0151/ — this address never moves; the anchor /records/#eng-0151 keeps resolving
36.62 Whmean 171.44 W − 10.1 W idle floor (box-idle-all-empty) = 161.34 W attributable
window 2026-08-17T10:07:10Z → 2026-08-17T10:19:59Z · 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 Vulkan d204800 (cfg-0105), covering both the prefill row (run-0381) and decode row (run-0382) from the single llama-bench invocation. Session window taken from the launch timestamp embedded in the run-0382 record's own comment (10:07:10Z launch) cross-checked against the ~/bench-results/depth-backfill-20260817/ornith-35b/vk-d204800.json file mtime (10:19:59Z, used as the end boundary); reconstructed, not runmeta-recorded (same gap as eng-0147..eng-0150). Counter interpolated from raw ~10s-resolution HA history samples: start 25.208454 kWh (10:07:10Z), end 25.245075 kWh (10:19:59Z). Slightly lower mean draw than the ROCm arms (171.44 W vs ~179 W) despite Vulkan's higher decode throughput at this depth (26.53 vs 22.23 t/s, per run-0382's comment) -- a same-model cross-backend efficiency contrast worth surfacing per protocol.json's compute-surfaces comparison rule.
Wh, never joules — read as a cumulative-counter difference across the window; delta over the idle baseline is the only figure that means anything