Salvato in:
Dettagli Bibliografici
Autori principali: Newkirk, Alex C., Fernandez, Jared, Koomey, Jonathan, Latif, Imran, Strubell, Emma, Shehabi, Arman, Samaras, Constantine
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:https://arxiv.org/abs/2506.14551
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866911295778848768
author Newkirk, Alex C.
Fernandez, Jared
Koomey, Jonathan
Latif, Imran
Strubell, Emma
Shehabi, Arman
Samaras, Constantine
author_facet Newkirk, Alex C.
Fernandez, Jared
Koomey, Jonathan
Latif, Imran
Strubell, Emma
Shehabi, Arman
Samaras, Constantine
contents As AI's energy demand continues to grow, it is critical to enhance the understanding of characteristics of this demand, to improve grid infrastructure planning and environmental assessment. By combining empirical measurements from Brookhaven National Laboratory during AI training on 8-GPU H100 systems with open-source benchmarking data, we develop statistical models relating computational intensity to node-level power consumption. We measure the gap between manufacturer-rated thermal design power (TDP) and actual power demand during AI training. Our analysis reveals that even computationally intensive workloads operate at only 76% of the 10.2 kW TDP rating. Our architecture-specific model, calibrated to floating-point operations, predicts energy consumption with 11.4% mean absolute percentage error, significantly outperforming TDP-based approaches (27-37% error). We identified distinct power signatures between transformer and CNN architectures, with transformers showing characteristic fluctuations that may impact grid stability.
format Preprint
id arxiv_https___arxiv_org_abs_2506_14551
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Empirically-Calibrated H100 Node Power Models for Reducing Uncertainty in AI Training Energy Estimation
Newkirk, Alex C.
Fernandez, Jared
Koomey, Jonathan
Latif, Imran
Strubell, Emma
Shehabi, Arman
Samaras, Constantine
Hardware Architecture
As AI's energy demand continues to grow, it is critical to enhance the understanding of characteristics of this demand, to improve grid infrastructure planning and environmental assessment. By combining empirical measurements from Brookhaven National Laboratory during AI training on 8-GPU H100 systems with open-source benchmarking data, we develop statistical models relating computational intensity to node-level power consumption. We measure the gap between manufacturer-rated thermal design power (TDP) and actual power demand during AI training. Our analysis reveals that even computationally intensive workloads operate at only 76% of the 10.2 kW TDP rating. Our architecture-specific model, calibrated to floating-point operations, predicts energy consumption with 11.4% mean absolute percentage error, significantly outperforming TDP-based approaches (27-37% error). We identified distinct power signatures between transformer and CNN architectures, with transformers showing characteristic fluctuations that may impact grid stability.
title Empirically-Calibrated H100 Node Power Models for Reducing Uncertainty in AI Training Energy Estimation
topic Hardware Architecture
url https://arxiv.org/abs/2506.14551