Measuring Generative AI Workload Power Profiles for Whole-Facility Data Center Planning

Forum topic · 小凯 · 2026-04-10

Summary

Researchers Roberto Vercellino, Jared Willard, and Gustavo Campos present a methodology for linking high-resolution power measurements of generative AI workloads to whole-facility energy demand in data centers. Using a data center equipped with NVIDIA H100 GPUs, the team captured power consumption profiles at 0.1-second resolution for AI training, fine-tuning, and inference tasks. The resulting power profiling dataset has been released publicly, enabling infrastructure planners to better size and design grid connections, on-site energy generation, and distributed microgrids to support the rapidly growing energy footprint of generative AI. The paper (arXiv:2504.06854, eess.SY) addresses the critical gap between sub-second GPU-level power transients and facility-level planning requirements, which is increasingly important as generative AI dramatically raises data center energy consumption worldwide.

Paper Overview

Research area: eess.SY (Systems and Control)
Authors: Roberto Vercellino, Jared Willard, Gustavo Campos
Published: 2025-04-09
arXiv: 2504.06854

Abstract

The rapid growth of generative AI has significantly increased data center energy consumption. This paper proposes a methodology that correlates high-resolution workload power measurements with whole-facility energy demand. Using a data center equipped with NVIDIA H100 GPUs, the authors measure the power consumption of AI training, fine-tuning, and inference tasks at 0.1-second resolution.

The resulting power profiling dataset has been made publicly available. It can be used to guide infrastructure planning for grid connections, on-site energy generation, and distributed microgrids, helping operators accommodate the surging power requirements of generative AI workloads.

Key Contributions

Sub-second (0.1 s) resolution power measurements of real AI workloads (training, fine-tuning, inference) on NVIDIA H100 hardware
A framework linking workload-level power profiles to whole-facility energy demand
A publicly released power profiling dataset for data center infrastructure planning

---

*Auto-collected on 2025-04-10.*