eng-0226
window recorded
citable URL: https://halobench.com/records/eng-0226/ — this address never moves; the anchor /records/#eng-0226 keeps resolving
62.10 Whmean 157.33 W − 10.1 W idle floor (box-idle-all-empty) = 147.23 W attributable
window 2026-08-22T16:56:54Z → 2026-08-22T17:20:35Z · 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 window) · no contention
per task 12.4203 Wh · 0.3763 p at 30.3 p/kWh
covers run-0553
LLaDA2.2-flash tau2 airline 5-task SCREEN (run-0553): block-diffusion on the headbouyJB/diffuse-cpp fork (GPU-resident KV cache + full Levenshtein editing), GPU-accelerated on aihydra (gfx1151), pinned Haiku-4.5 simulator, all 5 tasks passed. Counter interpolated from raw ~10s HA history: start 32.737790577 kWh, end 32.799891990 kWh, delta 0.062101413 kWh = 62.1014 Wh. wh_per_task is Wh per CORRECT answer: wh_total / tasks_passed (5 of 5). Whole-session energy includes simulator wait time end to end, matching the site's headline denominator. Retroactive join (the run was not harness-wrapped) within the ~10-day HA retention window; the window was verified contention-free — a dedicated diffuse-server was the only GPU consumer and the counter is smooth and monotonic. This is a 5-task screen, not the full 26-task arm: at 12.42 Wh per correct answer it is ~2.3x Ling-3.0-flash's 5.32 Wh per correct on its full run, an early cross-paradigm indicator (diffusion is compute-heavy per committed token), not a settled comparison (different N and task mix).
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
claims clm-0102
model pages llada22-flash