DIGITAL LIFE

New guide offers framework to measure AI's energy consumption

Advanced computing and data centers are massive energy consumers. In fact, a recent report from Lawrence Berkeley National Laboratory found that U.S. data centers consumed about 4.4% of total electricity in 2023 and projects them to potentially triple consumption by 2028. Widespread deployment of artificial intelligence (AI) is a driving factor of this rise in energy consumption and is already having an economic impact as electricity costs rise for consumers.

Researchers at the National Renewable Energy Laboratory (NREL) are shining a light on this energetic cost of computing. Along with investigating methods to make computing more energy efficient, researchers are working to disseminate tools and insights beyond the research sphere to industry software professionals.

A recent result of these efforts is "A Beginner's Guide to Power and Energy Measurement and Estimation," a new NREL report developed in partnership with Intel that outlines key considerations for machine learning developers and practitioners seeking to use energy measurement tools and interpret energy estimates.

"With AI playing a growing role in both research and industry, its increasing impact on energy consumption has become a shared challenge we can tackle together," said Hilary Egan, data scientist and lead NREL author on the report. "Through this guide, we wanted to provide AI professionals with an introduction to energy estimation that opens the door to more sustainable decision-making in computing."

Taking AIm: Partnering to address AI's energy use challenge...NREL's commitment to reducing computing's—and AI's—energy consumption inspired the launch of the Joint Institute for Strategic Energy Analysis (JISEA) Green Computing Catalyzer in 2022. The Green Computing Catalyzer is one arm of JISEA's Catalyzer Program exploring strategic areas for investment in future energy opportunities and brings together researchers, universities, and industry partners to analyze pathways to reduce computing's energy impact.

Over its tenure, the Green Computing Catalyzer has supported efforts to quantify and catalog the energetic costs of machine learning and NREL's advanced computing systems, with the goal of providing a means for more transparency and accountability in computing. These efforts caught the attention of Intel, which leads its own collection of Responsible AI initiatives.

Together, the Green Computing Catalyzer and Intel developed "A Beginner's Guide to Power and Energy Measurement and Estimation," aiming to equip developers with the skills they need to conduct productive energy measurement in their computing. This allows developers and their companies to make informed decisions about the sustainability of their systems.

"Sustainability has been a longstanding priority at Intel, both before the advent of AI and now," said Ronak Singhal, senior fellow in the Datacenter and AI group at Intel. "NREL's contributions were crucial in bringing to life our shared publication, which equips developers with the skills to make intelligent measurement decisions—a vital first step on the road towards sustainability in AI."

Shaping a measurement framework for AI's energy usage...While some AI/machine-learning developers and cloud companies are beginning to incorporate energy considerations into the development of their various models, there is no uniform standard framework for measuring energy use across all computational levels. The new guide serves as a helpful resource to make energy efficiency tools and procedures more accessible, while establishing a framework for measuring energy at the system, job, application, and code levels.

The report offers a roadmap for determining computing's energy usage from both hardware and software perspectives and discusses the challenges of interpreting these measurements as useful estimates. It also contains practical tips and real-world scenarios that illustrate the application of different energy considerations that can be used across the computing industry.

Practitioners using the guide will start at the beginning of the energy project measurement workflow: determining the key questions to answer. The guide then walks through how to determine the appropriate and feasible measurement tools to gather the data that then must be analyzed and interpreted. Finally, the roadmap helps practitioners and their companies determine the sufficiency of the analysis, which may lead to refinement of questions and measurements.

Overall, the report provides measurement, analysis, and interpretation guidance from the system level—how a collection of workloads affects energy usage—to the code level—quantifying the energy consumed by parts of a job or application.

"The Catalyzer Program addresses critical, intersectional energy challenges," said NREL's Kristin Wegner Guilfoyle, Catalyzer Program lead. "This roadmap helps achieve that goal by connecting the progress made by energy-efficient computing researchers with the innovations driven by software developers and practitioners."

Provided by National Renewable Energy Laboratory