Workshop: EESP 2026: 3rd International Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices
Authors: Kevin Menear, Kumaraguru Prabakar, and Deepthi Vaidhynathan (National Laboratory of the Rockies)
Abstract: Different instruments report different power for the same AI workload. We show this on production H100 racks by recording the same events at three concurrent tiers: one-minute breakers, 100 ms per-GPU telemetry, and 10 kHz voltage and current on every rack feed. Disagreement between tiers shows what each instrument resolves. Coarse tiers return wrong numbers with no sign they are wrong: a training job pulses 5.75 times per second, and device telemetry, sampling ten times per second, reports 4.18. Only the feed record shows that a resumed job’s first pulse lands at full amplitude, and that an inference-serving instance under burst admission swings its whole rack from idle to peak twelve times per second. A resolution ladder states which tier resolves each phenomenon and how the others fail. For this load, ramp rates, transient magnitudes, and the coherence of aggregated units are sub-cycle quantities.
Back to EESP 2026: 3rd International Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices Archive Listing
Back to Full Workshop Archive Listing