Name the electrical boundary before reporting energy

Power is the rate of energy transfer. A constant reading of one watt maintained for one second corresponds to one joule; a brief high-power burst may consume less energy than a long low-power wait. Average power only becomes an energy estimate when multiplied by the duration over which that average applies. A peak reading, a power limit, and a thermal design value are not interchangeable with the measured average. Preserve the original timestamps rather than multiplying whichever dashboard number is most convenient.

Watt-hours express energy, despite containing the word watt. One hour contains 3,600 seconds, so one watt-hour equals 3,600 joules, and one kilowatt-hour equals 3,600,000 joules. The conversion does not depend on a hardware model. Using kilowatts with seconds without an additional conversion introduces a factor error. Write intermediate units alongside numbers; dimensional cancellation often exposes a mistake before a plausible-looking efficiency ratio hides it.

Draw nested boxes for accelerator circuitry, the server, and the facility. A whole-server wall measurement normally already includes the accelerator and power-supply losses within that server boundary. Adding a GPU-board counter to it counts part of the electricity twice. To construct a node estimate from components, require disjoint component coverage, synchronized intervals, and a declared treatment of conversion losses. A quantity labeled host power must say whether it excludes the accelerator rather than relying on an ambiguous label.

A measurement also has a temporal boundary. Starting before initialization includes model loading and cache preparation; starting afterward measures another workload phase. Ending when an asynchronous kernel is launched omits the subsequent device execution. Define the completion event and account for all work relevant to the question, including retries or cleanup if they are within scope. Report excluded phases separately when their frequency could change the comparison between implementations.

Green Algorithms provides an explicit estimation framework using runtime, hardware assumptions, utilization, and facility or grid inputs. Such a model can make missing measurements visible and support sensitivity analysis. It does not turn an assumed component power into an observed electrical trace. Store whether each input was metered, inferred, or assigned from a default. A reproducible estimate with an uncertainty range is more informative than a precise decimal whose measurement provenance is unavailable.

Shared infrastructure requires an additional accounting decision. A node counter covers all concurrent work on that node, not automatically one process. Assigning a fraction by time, reserved capacity, or measured activity is an allocation policy whose totals should reconcile to the parent account. Incremental energy relative to an idle baseline answers a different question from gross electricity during execution. Keep both the parent measurement and the allocation rule so another reader can reconstruct the reported job number.