Measured on an 8-node HPC cluster by DCIMate: wall power (IPMI DCMI), CPU power (RAPL), die temperature, fan RPM, and a Kubernetes A/B of 30 llama.cpp jobs.

Modelled from those measurements: noise (fan law on the measured RPM), PUE and CRAH duty, and 175 W per switch. No acoustic or facility telemetry exists.

Switching nodes off is the twin's own control, not a dcimclaw feature: it parks nodes instead, and never measured a powered-down one. The 15 W standby is an assumption, though the saving holds from 5 W to 40 W.

One cabinet is that instrumented cluster plus a ToR switch; the hall is 10 of them plus 2 network racks.

Validated against the measured A/B: job durations within 5%, per-node wattage to about 1% of cluster power, and energy per job within 9%.

Weakest link: fan speed. The two measurement campaigns disagree on the chassis fan curve by up to 1.9x, so read RPM and dB(A) as direction and rough magnitude, not as calibrated values.