Stony Brook receives $1 million from DARPA to trace and reduce AI energy use

Stony Brook researchers have received a two-year, $1 million DARPA award to develop tools that show where machine-learning models consume energy. The goal is to help developers cut consumption while preserving the accuracy their applications need—a research effort aimed at making energy use a more precise part of software and hardware decisions.

Stony Brook described the award on October 20, 2026. The project, “Optimizing Machine Learning Models Using Accurate Energy Prediction,” is funded through the Defense Advanced Research Projects Agency’s Mapping Machine Learning to Physics program. It targets a gap between estimating a model’s total energy consumption and understanding enough detail to decide how to reduce it.

Following energy through a model

The proposed approach starts by breaking a machine-learning model into its basic components. Researchers will represent its execution as a structured graph: a map of those components that captures both their individual energy consumption and how their contributions combine into the model’s overall energy use.

That detail is central to the research. Existing prediction methods often give broad estimates, according to the university. They can miss how parts of a model interact with one another, with techniques used to optimize them, and with the hardware running them. The team wants its accounting to capture those relationships rather than stop at an overall estimate.

The intended result has two uses: better predictions of energy consumption and a clearer diagnosis of where energy is being spent. Pinpointing those locations would help developers identify bottlenecks and direct their efforts toward the parts of a model that need attention.

“They are very helpful for getting a broad estimate, but they are not precise enough when you need to decide things like which model to deploy or which hardware to run it on.”

Aruna Balasubramanian, project principal investigator, on existing energy-prediction approaches

Choosing for the task and the device

The practical goal is to help developers compare options for a particular task and hardware platform. Rather than treating energy use separately from application requirements, the planned tools would show tradeoffs between energy and performance. The university identifies several decisions the tools could support:

• Model selection: determine which model best suits the task while meeting the accuracy the application requires.
• Optimization choices: compare techniques for reducing consumption in the context of the model and hardware being used.
• Hardware selection: assess where to run the model, including on resource-constrained devices such as drones.

The consequences differ by setting. For services that continually process requests, energy consumption adds to operating costs. On phones, sensors and autonomous systems, it can limit the time a device operates before recharging. Those are the deployment constraints motivating the work, not measured savings delivered by the project.

A two-year research deliverable

Aruna Balasubramanian leads the Stony Brook team as principal investigator, with computer science faculty members Niranjan Balasubramanian and Anshul Gandhi as co-principal investigators. Xiaorui Wang of Ohio State University is also a principal investigator on the project.

The team plans to develop the tools by the end of the project; this is not a software release or a completed finding about energy savings. DARPA’s broader program aims to improve the use of limited power resources. Stony Brook says the program’s expected open-source research outputs are intended to support energy-aware applications running on local devices, including in defense industrial settings.