FURI | Spring 2026

Analyzing GPU Power Behavior in AI Workloads for Energy-Efficient Data Centers

This project focuses on analyzing GPU power behavior in AI workloads using large-scale data from the MIT Supercloud dataset (~2TB). Each job is represented as high-frequency time-series data, capturing how GPU power changes over time. A key challenge is that the dataset is unlabeled, meaning workload types such as training, fine-tuning, and inference are not directly provided.

To address this, we are developing a data analysis approach using Python to process and visualize GPU power traces. We are extracting statistical features such as duty cycle, peak-to-idle ratio, and variability to identify patterns that may distinguish different workload types. We are also applying frequency-domain analysis (FFT) to explore whether seemingly random power spikes follow underlying periodic behavior.

The goal of this work is to better understand transient GPU power usage and its impact on energy consumption. This research aims to support the development of more efficient, stable, and energy-aware data center operations as AI systems continue to scale.