Modeled US volume-server idle/maximum power dynamic range: 0.44 minimum by 2020 from measured SERT data, 0.28 in the Best Practices scenario, with an explicit ~20-25% idle-power latency floor

MODELED, and the distinction between measured and modeled matters here. The 2016 DOE/BL report states that minimal documentation exists on the average dynamic range of the US installed server stock, then bounds it between a 'Maximum DR' trend (worst performers, the 0.67 assumption carried over from the 2007 report, sloping linearly to 0.44 in 2020) and a 'Minimum DR' trend derived from SPEC power database values. The 0.44 figure is the SERT-database-derived measured anchor ('servers performing with an average dynamic range of ~0.44'); the report assumes it as the floor all volume servers achieve by 2020. The current-trends average is a 90/10 mix of the max and min DR trends; the Best Practices scenario assumes 50/50 penetration of the worst trend, giving 0.28 by 2020. Dynamic range is defined as the ratio of the lowest power level (idle) to maximum power, and is combined with average utilization assumptions to produce average server power.

THE GENERAL FORM OF THE H200 IDLE FINDING, and it says the 41.7% idle-to-loaded ratio recorded for the H200 node is NOT AN H200-QUIRK - an idle-to-maximum ratio near 0.44 for volume servers is the assumption a DOE/BL model built on measured SERT data actually uses. Note the precise framing: dynamic range is idle divided by MAXIMUM, whereas the repo's H200 figure is idle divided by the MEASURED LOADED plateau, so the two are close but not identical quantities and must not be quoted as the same statistic. THE 20-25% FLOOR IS THE MOST USEFUL PART: the report states that getting below 20-25% of maximum power at idle 'can require powering down specific functionality that may increase idle-to-active wake-up latency and result in delayed response times'. So there is a measured reason not to chase idle power toward zero, and any 'power-proportional fleet' plan should price first-token latency. Accelerator-specific idle ratios are NOT derivable from this record: it covers volume x86 servers, and no MI300X, L40S, A100 or TPU idle ratio is recorded anywhere in this task - the underlying SERT and SPEC datasets are the place to look next. METRIC CORRECTION 2026-10-05 (schema authority pass, agent w5-schema). Value UNCHANGED; `metric` moved quality -> dimensionless_ratio. Both members were live in this record's case and the enum forced a choice between two wrong answers: `quality` implies a score and this is a ratio, and there was no dimensionless member at all. This record is one of the five named in gotcha bench-metric-enum-hides-a-dimensionless-ratio, whose root_cause records that the `quality` catch-all was carrying accuracy deltas, PUE ratios and power ratios at once. METRIC CORRECTION 2026-10-05 (schema authority pass, agent w5-schema). Value UNCHANGED; `metric` moved quality -> dimensionless_ratio. idle power / maximum power is a ratio with no faster/slower reading, so it is not a speedup. Previously filed as `quality`, the catch-all that gotcha bench-metric-enum-hides-a-dimensionless-ratio names as the reason `quality` cannot be trusted as a semantic filter.