Energy and carbon accounting

Power is an instantaneous rate; energy accumulates over time. Device telemetry may measure only an accelerator, while wall measurements include processors, memory, storage, fans, and power conversion. Record which components are included, sampling interval, meter accuracy, and treatment of idle power. Allocating a shared host's energy to one job requires a stated method. A nameplate power limit is not a measurement of average consumption.

If measured energy covers IT equipment only, a datacenter overhead factor can estimate facility electricity. Power usage effectiveness is facility energy divided by IT energy over a defined interval. It should not be applied again to a measurement that already includes facility overhead. For short runs, compilation, data loading, and idle allocation can matter substantially. Count failed experiments and tuning when the claim concerns a research project's total footprint rather than one successful final run.

Operational emissions estimates multiply electricity use by an appropriate carbon-intensity factor. State whether the factor is average or marginal, location- and time-specific or coarse, and which gases its carbon-dioxide-equivalent units include. These factors answer different questions and should not be mixed silently. A contractual renewable-energy claim is not automatically evidence that the physical grid electricity used during the run had zero emissions.

Hardware manufacture and disposal also have emissions. Allocating a share of embodied emissions can use an explicit lifetime and reserved-resource share, but these are accounting assumptions with uncertainty. The Software Carbon Intensity specification includes operational and allocated embodied components per functional unit. Keep the raw totals alongside the intensity. A lower per-request estimate can coexist with rising total emissions when request volume grows, so efficiency claims should not be presented as total-impact claims.

Choose a functional unit that preserves utility: a training run reaching a fixed quality target, or a completed inference task meeting quality and latency requirements. Tokens alone may reward verbose outputs; accelerator-hours alone ignore power. Report quality, latency, energy, operational emissions assumptions, and embodied allocation separately. This lets readers recompute a comparison with a different electricity factor instead of treating one composite score as immutable.

Potential reductions include better utilization, less redundant computation, model reuse, and scheduling flexible work in lower-intensity periods. They have constraints: moving jobs can add data transfers, delay useful results, or alter capacity needs. Evaluate the complete change instead of assuming that relocation always helps. Give sensitivity ranges for uncertain power, overhead, grid intensity, and hardware lifetime, and distinguish measurement uncertainty from choices about the accounting boundary.