Collaborative Research: EAGER: CET: Exploring The Risks and Rewards of Large Language Models in Enabling Energy-Efficient Data Center Software Infrastructure

This EArly-concept Grants for Exploratory Research (EAGER) award is made in response to Dear Colleague Letter 23-109, as part of the NSF-wide Clean Energy Technology initiative. This project works to discover more efficient ways of using energy to mitigate global climate change and to promote U.S. national security through greater energy independence.

The project targets the optimization of energy consumption in data centers, which currently consume ~ 2% of the U.S.'s energy use and is projected to grow substantially. The project’s objectives are to develop techniques to improve the energy efficiency of software, so that less energy is spent in data centers. The scientific innovation of the project is in the application of an emerging technology, Large Language Models (LLMs), to this task.

Additionally, the principal investigators will use a systems thinking conceptual framework to identify patterns and understand interconnections needed to develop new educational material related to this topic. This material will be presented at a workshop in the second year of the project and made freely available online. More generally, the possible contribution is to catalyze a shift toward effective energy-efficient computing, both in software engineering practices and in STEM education.

With funding from an EAGER award through the NSF-wide Clean Energy Technology initiative, this project investigates approaches to improve the energy efficiency of software and fill two gaps in the current energy-aware toolkit - energy-aware tools to support software implementation, especially for data center software, and energy-focused educational material for university students and workforce development. Toward the first goal, the project studies the application of LLMs in developing energy-efficient software solutions for data centers.

The principal investigators evaluate the hypothesis that LLMs can significantly improve software energy efficiency with targeted prompting and feedback from modeled and real energy data. Toward the second goal, the project assesses the effectiveness of systems thinking as a conceptual framework for energy efficiency. They hypothesize that systems thinking, with its emphasis on holistic reasoning across levels of abstraction and timescales, is a better basis than the prior conceptual frameworks.

To validate these hypotheses, the project is structured around three thrusts: measuring the efficacy of LLMs in fostering energy-efficient programming, creating pedagogical materials, and disseminating our insights through a workshop. The research specifically contributes (1) a framework for automated energy-optimization of software, including a catalog of energy-efficiency LLM prompts combined with a hierarchy of energy measures; (2) empirical data on the effectiveness of LLMs in developing energy-efficient software for use in data centers; and (3) educational materials on energy efficiency for software in data centers, including the first incorporation of LLMs and of systems thinking into this line of pedagogy.

If successful, the project will advance the state of the art in energy-aware software engineering and pedagogy.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.