Optimizing AI energy consumption throughout its entire lifecycle
Funded with $1.83M from DARPA, University of Michigan Engineering leads a project that aims to cut AI’s energy needs by up to half.
Artificial Intelligence
Explore the forefront of AI
at Michigan Engineering
RESEARCH TO WATCH
What’s the study
Data center energy consumption is expected to more than double by 2030 from roughly 420 terawatt-hours in 2024, with much of the anticipated growth driven by AI. But more efficient software design could cut AI energy use by up to half, according to University of Michigan Engineering researchers teaming up with colleagues at the University of Illinois Urbana-Champaign (UIUC) and defense contractor BAE Systems.
Rather than relying on expensive, time-consuming trial-and-error testing across thousands of hardware and software configurations, the team’s proposed system will enable developers to estimate energy usage across every stage of an AI workflow—from data processing and model training to real-world deployment. Over the course of the project, the researchers will build high-precision energy measurement tools that combine physical sensors with software analysis, releasing them as open-source documentation and software. The team calls the new mathematical framework underpinning these tools “energy calculus.”
Why it matters
While computer chips are getting faster, misalignments between software and hardware waste massive amounts of electricity—up to 30% during AI model training alone. More energy-efficient AI brings down electricity demand, reducing the cost of AI and extending battery life for AI-enabled devices. This is especially critical for national defense and military technology, where field equipment and remote computing systems must run complex AI algorithms under strict battery and power limits.
How much and for how long
The $1.83 million contract from the Defense Advanced Research Projects Agency (DARPA) will fund the two-year project, titled, “E-CALC: Energy Calculus for ML Lifecycle Optimization.”
The initiative builds on previous efforts by the team lead. For AI model training, the Zeus energy optimization framework automatically tunes settings like batch size and power limits to minimize energy use. Exploring the energy cost of running AI models, the ML.ENERGY benchmark acts as a public leaderboard that measures and compares energy consumption across different open-weight AI models and open-source software frameworks.
Who’s involved
Mosharaf Chowdhury is an associate professor of computer science and engineering at the University of Michigan and principal investigator for E-CALC. He leads ML.ENERGY and Zeus projects.
“AI is becoming part of everyday life, business, science and national security, but it requires enormous and rapidly growing amounts of electricity. If that energy use continues to rise unchecked, it could increase costs, strain power infrastructure, and make advanced AI harder to deploy. This project tackles a practical question: how can we get widespread benefits of AI with minimum energy?” said Chowdhury.
The two co-principal investigators are Nishil Talati, an assistant professor of electrical and computer engineering at the University of Illinois Urbana-Champaign specializing in hardware-software co-design and execution bottleneck characterization, and David Nicholson, a principal research engineer at BAE Systems who leads internal research and development on edge AI technologies for defense applications.
Approved for public release, distribution unlimited. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Defense Advanced Research Projects Agency (DARPA).