Empirical Measurements of AI Training Power Demand on a GPU-Accelerated Node

TL;DR

This paper empirically measures the power consumption of an 8-GPU NVIDIA H100 node during AI training, revealing insights into energy use and efficiency improvements for data center planning.

Contribution

It provides the first detailed empirical power measurements of an NVIDIA H100 GPU node during AI training, highlighting energy consumption patterns and efficiency gains with batch size adjustments.

Findings

Maximum power draw was approximately 8.4 kW, below the rated 10.2 kW.

Increasing batch size from 512 to 4096 reduces energy consumption by a factor of 4.

Power consumption patterns inform data center capacity planning and energy estimates.

Abstract

The expansion of artificial intelligence (AI) applications has driven substantial investment in computational infrastructure, especially by cloud computing providers. Quantifying the energy footprint of this infrastructure requires models parameterized by the power demand of AI hardware during training. We empirically measured the instantaneous power draw of an 8-GPU NVIDIA H100 HGX node during the training of open-source image classifier (ResNet) and large-language models (Llama2-13b). The maximum observed power draw was approximately 8.4 kW, 18% lower than the manufacturer-rated 10.2 kW, even with GPUs near full utilization. Holding model architecture constant, increasing batch size from 512 to 4096 images for ResNet reduced total training energy consumption by a factor of 4. These findings can inform capacity planning for data center operators and energy use estimates by researchers. Future work will investigate the impact of cooling technology and carbon-aware scheduling on AI workload energy consumption.