Sustainability Reporting for Accelerated Computing

Reporting energy use for accelerated computing credibly requires more than PUE, which understates the picture in ways that matter for AI workloads specifically.

Why PUE understates it

PUE measures facility overhead against IT load. It says nothing about whether the IT load is doing useful work. A cluster at forty percent utilisation and a cluster at eighty can report the same PUE while producing very different amounts of output for the same energy, and the facility metric will look identical.

For AI workloads this gap is wide, because idle accelerators still draw a significant fraction of peak power. Reporting only PUE therefore describes the efficiency of the building while saying nothing about the efficiency of the computing.

What to report alongside it

Energy per unit of useful work is the number that actually reflects efficiency: per training run, per million tokens served, or per experiment, depending on what the estate does. It is harder to produce and it is the one that improves when utilisation improves, which is where the real gains are.

Utilisation itself should be reported, with the caveat covered elsewhere in this library that GPU utilisation percentage is a poor signal on its own. Achieved throughput is more honest than resident-kernel time.

Cooling choice belongs in the report

In this climate cooling is a substantial fraction of facility power, so the cooling strategy is a reportable decision rather than an implementation detail. Warm water cooling that avoids a chiller changes facility energy materially, and it is a legitimate and verifiable claim in a way that most efficiency statements are not.

Be honest about boundaries

State what is included. Embodied carbon in hardware, network equipment outside the hall, and energy consumed by development work on desk-side systems are commonly excluded without being declared. A narrower boundary honestly stated is more credible than a broad one quietly drawn.