Empirically-calibrated H100 node power models for accurate AI training energy estimation (Newkirk et al., Environmental Research: Energy)

DOI 10.1088/2753-3751/ae2486. Measured 8-GPU H100 node power during real AI training at Brookhaven National Laboratory and calibrated statistical models to FLOPS intensity. Headline: even compute-intensive workloads ran at only 76% of the 10.2 kW node TDP rating. The architecture-specific model predicts energy with 11.4% mean absolute percentage error versus 27-37% error for TDP-based estimates. Also finds distinct power signatures for transformer vs convolutional architectures. The TDP-vs-actual gap is the general lesson for anyone computing cost from nameplate watts.