New DARPA-funded project aims to cut AI energy use by up to half

A two-year, $1.83 million project led by University of Michigan Engineering researchers aims to cut the energy consumed by artificial intelligence by up to half, by giving developers a way to estimate power use across every stage of an AI workflow before expensive hardware is ever switched on. The contract comes from the Defense Advanced Research Projects Agency, or DARPA, and the effort is named E-CALC: Energy Calculus for ML Lifecycle Optimization.

The motivation is scale. 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. According to the University of Michigan researchers, who are teaming up with colleagues at the University of Illinois and the defense contractor BAE Systems, more efficient software design could cut AI energy use by up to half.

The core problem the team plans to tackle is a mismatch between software and hardware. While computer chips keep getting faster, misalignments between the code that runs on them and the silicon beneath can waste massive amounts of electricity—up to 30 percent during AI model training alone, the researchers note. Today, finding energy-efficient configurations typically relies on trial-and-error testing across thousands of hardware and software combinations, a process that is both expensive and time-consuming.

E-CALC proposes a different approach. Rather than measuring everything by brute force, the team will build a mathematical framework it calls “energy calculus,” designed to let developers estimate energy usage across every stage of an AI workflow—from data processing and model training to real-world deployment. The project is a plan, not a finished product: over its two-year course, the researchers will build high-precision energy measurement tools that combine physical sensors with software analysis, and they intend to release them as open-source documentation and software.

The initiative builds on previous work 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. Separately, the ML.ENERGY benchmark acts as a public leaderboard that measures and compares the energy cost of running AI models across different open-weight AI models and open-source software frameworks.

Mosharaf Chowdhury, an associate professor of computer science and engineering at the University of Michigan, is the principal investigator for E-CALC and leads both the 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,” Chowdhury said. “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?”

Two co-principal investigators join the effort. Nishil Talati is an assistant professor of electrical and computer engineering at the University of Illinois who specializes in hardware-software co-design and execution bottleneck characterization. David Nicholson is a principal research engineer at BAE Systems who leads internal research and development on edge AI technologies for defense applications.

The defense angle is a significant part of the project’s rationale. More energy-efficient AI brings down electricity demand, reducing the cost of AI and extending battery life for AI-enabled devices. That 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.

Because E-CALC is a newly funded research program, its outcomes remain to be demonstrated; the up-to-half energy savings figure reflects the project’s stated goal rather than a measured result. What the funding does establish is a concrete research agenda: open-source measurement tools, a formal calculus for reasoning about energy across the AI lifecycle, and a collaboration spanning academia and defense industry. If the tools perform as intended, they could give developers a way to design for efficiency from the start—rather than discovering wasted watts after deployment.

The material has been approved for public release, and the researchers note that any opinions, findings, conclusions or recommendations expressed in the project do not necessarily reflect the views of DARPA.

Subject of Research: Energy-efficient AI software and hardware optimization across the machine learning lifecycle

Article Title: Optimizing AI energy consumption throughout its entire lifecycle

Article References: Optimizing AI energy consumption throughout its entire lifecycle. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: artificial intelligence, energy consumption, data centers, DARPA, University of Michigan, machine learning, energy efficiency, model training, software-hardware co-design, open-source tools, defense technology, E-CALC