University of Michigan Engineering is leading a two-year project to estimate and optimise energy consumption across an AI workflow. The October 27 announcement describes a $1.83 million DARPA contract for E-CALC, short for Energy Calculus for ML Lifecycle Optimization.
Measure more than the final model call
The proposed system covers data processing, model training and deployment. Researchers plan to combine physical sensors with software analysis to produce high-precision energy measurements, then release documentation and software as open source.
The team calls the mathematical framework behind those tools energy calculus. Its intended role is to help developers estimate the effect of hardware and software choices without testing every possible combination through repeated trial and error.
A target rather than a completed saving
The university says more efficient software design could reduce AI energy use by up to half. That is a stated possibility motivating the research, not a reported result from the new two-year project.
The distinction matters for comparisons. Energy use depends on the workflow, hardware and settings being measured. An estimate for one stage should not automatically become a claim about the full lifecycle or a deployment running under different conditions.
Build on existing measurement work
Principal investigator Mosharaf Chowdhury leads the ML.ENERGY and Zeus projects. The announcement describes Zeus as tuning training settings such as batch size and power limits, while ML.ENERGY compares energy use across open-weight models and software frameworks.
Co-principal investigators Nishil Talati at the University of Illinois Urbana-Champaign and David Nicholson at BAE Systems bring hardware-software and edge-AI experience.
For developers following E-CALC, useful outputs would make the workload and measurement boundary explicit. Comparing energy, runtime and required model quality together would help establish whether an optimisation is appropriate for the system that will actually run it.