Energy measurement tools for AI training and inference

Energy reporting for AI workloads is being requested by more than one audience at once. Sustainability teams need scope 2 and scope 3 figures for CSRD and customer questionnaires, platform teams want efficiency per training run and per inference request, and procurement teams are asked to compare hardware or model providers on something other than marketing claims. The Green Software Foundation SCI specification defines the shape of a defensible answer: operational emissions from energy multiplied by grid carbon intensity, plus a share of embodied emissions, expressed as a rate per functional unit.

The measurement itself is awkward. Accelerator power can be sampled through vendor APIs, CPU and memory power through hardware counters or out-of-band interfaces, and facility overhead only through a PUE figure that is rarely available per job. Grid intensity is location-based and varies by hour, while many dashboards still apply an annual national average. Embodied emissions have to be amortised over hardware life and allocated by time and resource share. Short jobs make coarse sampling unreliable, and shared clusters make allocation a policy decision rather than a measurement.

The usual failure is a number without a method. A tool reports only GPU power, ignores cooling and networking, uses nameplate TDP instead of measured power, or mixes market-based renewable claims with location-based intensity, producing totals that cannot be compared across teams or years.

Evaluate measurement coverage, sampling interval, functional unit, factor sources and the uncertainty statement before dashboard features. If the methodology cannot be explained to an auditor in one page, the number will not survive review.