Predictive Energy Modeling for GPU Workloads: A Machine Learning Approach to Sustainable Data Centers
Date:
Invited talk at the University of Idaho Graduate Seminar on using machine learning to predict GPU energy consumption and improve data center workload scheduling. This work surrounds the broader topic of energy efficient datacenters.
Abstract
The rapid expansion of cloud computing and AI-driven applications has placed immense pressure on data center infrastructure, with GPU-intensive workloads contributing significantly to rising energy demands. Optimizing workload management for energy efficiency has become a critical challenge, particularly in addressing the high power consumption and cooling costs associated with GPU workloads.
This talk presents a machine learning-based approach to predicting GPU energy consumption, enabling intelligent workload scheduling to reduce overall power usage and cooling overhead. By analyzing the power characteristics of GPU workloads, predictive models have been developed to enhance workload distribution, ensuring efficient resource allocation while minimizing unnecessary energy expenditure.
Beyond predictive modeling, the talk explores broader research on energy-efficient workload management, including thermal-aware scheduling strategies, data placement optimization, and real-world workload simulations used to evaluate energy-saving techniques. The role of ML-driven workload distribution in improving sustainability is highlighted, demonstrating how advancements in machine learning and distributed computing can contribute to greener and more efficient data center operations.