Workload energy attribution
How physical GPU and host energy becomes a workload-level environmental record.
Abstract. The programme connects direct board power and cumulative energy counters with activity-based attribution and model validation. The objective is not merely to report watts, but to establish an auditable path from physical hardware to a workload, job or GPU-hour.
Task-normalized energy accounting; primary measurement campaign pending.
Direct telemetry establishes the physical anchor.
A reference A100 telemetry record exposes 409.692 W board power and 319,159,391 mJ of cumulative energy — 88.655 Wh at the observation point.
Instantaneous GPU board power in the reference record.
Hardware energy counter used as the integration anchor.
Cumulative energy converted to a more usable unit.
Energy counter differences provide an independent power signal.
Evidence status: the 409.692 W telemetry example is a working reference record. Source provenance must be attached before it is promoted to a formal Critical Load publication result.
The difficult step is allocation, not sensing.
Node and GPU energy is shared. A useful compute metric requires an explicit attribution rule.
Observe
Read GPU and host energy counters plus utilization and process activity.
Separate
Estimate idle and dynamic power over the same observation window.
Attribute
Allocate dynamic energy using workload activity and shared resource time.
Integrate
Convert power samples into energy for the job, pod, process or GPU-hour.
Record
Preserve source, estimator version, window and uncertainty in the evidence record.
The model must expose the allocation of shared idle power and the activity signal used for dynamic power.
Power estimators can be tested against a physical reference.
In the reference validation, estimator-side mean absolute error was materially lower than the local model.