Workshops Archive

Watts Behind the Weights: Characterizing Energy Consumption of the AI Training Lifecycle in HPC


Workshop: AI on HPC: Performance Engineering, Challenges and Opportunities

Authors: Alexandre Fernandes, Alícia Oliveira, Beatriz Cepa, Cláudia Brito, and Ricardo Macedo (INESC TEC, University of Minho)

Abstract: Modern AI training workloads have become a leading contributor to energy consumption in HPC infrastructures. Yet this consumption is treated as a monolithic block, attributed to execution alone. The training lifecycle includes preparation and post-execution stages, and execution itself comprises multiple phases whose energy consumption varies with the task, model, dataset, and infrastructure-related factors. This work closes this gap by characterizing energy consumption across the entire AI training lifecycle in HPC infrastructures, following a fine-grained approach that attributes energy to individual phases and events, rather than to training as a whole. We apply this methodology on a production supercomputer, assessing preparation and post-execution using distinct datasets and models, and characterizing execution over workloads with distinct resource requirements, namely the I/O-intensive ResNet-50 and the GPU-bound Qwen3-8B. We draw several lessons to help HPC users accurately attribute energy consumption across their workloads and make their training pipelines more energy-efficient.


Back to AI on HPC: Performance Engineering, Challenges and Opportunities Archive Listing

Back to Full Workshop Archive Listing