MIT and IBM Watson AI Lab Develop Faster Tool to Predict AI Power Consumption for Sustainability

Researchers at MIT and the MIT-IBM Watson AI Lab have created a rapid prediction tool that estimates the power consumption of AI workloads across various processor chips. Presented at the IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS), this tool aims to address the growing energy demands of data centers, which could consume up to 12% of total U.S. electricity by 2028. Traditional methods for estimating power use rely on slow, time-consuming simulations, but this new approach provides a much faster alternative, enabling more efficient energy management.

The tool works by analyzing the specific characteristics of AI tasks and matching them to the power profiles of different hardware, such as CPUs, GPUs, and custom accelerators. This allows data center operators to quickly identify which workloads are most energy-intensive and optimize their hardware allocation accordingly. By improving energy efficiency, the tool not only reduces operational costs but also supports sustainability goals in an era of rapidly expanding AI adoption. The research underscores a critical shift toward making AI infrastructure more environmentally responsible.