Decision Yield: Useful Decisions per Joule or kWh

The formula

DY(Mx) = N_pass / E(Mx)

N_pass is the number of tasks that pass the quality gate, and E(Mx) is the energy at meter point Mx over the same period. It is reported in decisions per joule, and also per kWh.

Quality-gated, or it does not count

The task set is versioned and defined for each class of machine, such as language models, vision or robot navigation. A quality gate is stated, only passing tasks count, and accuracy is reported beside the ratio. This follows MLPerf, which reports samples per joule only for results that meet the benchmark's quality target (Tschand et al., MLPerf Power).

Energy per query, turned around

"Energy per query" figures are the same idea inverted: energy divided by answers. They only compare fairly when the boundary, the task and the quality bar match. See how to measure AI inference energy.

Questions

Why not count every request?

Because a system could raise its score by answering badly. Counting only passing tasks keeps the ratio honest.

Sources

Tschand et al., MLPerf Power

Keep reading

Energy per decision and Landauer GapFree calculator: turn measured energy and passing tasks into energy per decision and Decision Yield, and compare an operation with the Landauer limit.How to measure AI inference energyHow to measure the energy of AI inference properly: task set, quality gate, meter point, analyzer, ranging, 60-second windows, closure check and reporting.AI energy benchmarks comparedMLPerf Power, the AI Energy Score and the ML.ENERGY benchmark compared: what each measures, where its boundary sits, and how the Joule Ledger relates.Overhead Ratio (PUE)The JL-26 Overhead Ratio, facility energy over system energy: equal to PUE for data centers, calculated from energy over a period per ISO/IEC 30134-2:2026.

Take the next step.

Decision Yield, the JL-26 ratio of quality-gated decisions per joule at a stated meter point: formula, task sets, quality gates and MLPerf.

Energy per decision calculator Get the textbook