Project Launched to Cut Artificial Intelligence Energy Consumption in Half
Led by the University of Michigan and supported by a $1.83 million grant from DARPA, a team is developing an energy accounting framework aimed at reducing the energy demand of artificial intelligence workflows by up to 50 percent.
The University of Michigan College of Engineering is leading a two-year research project aimed at reducing the energy consumption of artificial intelligence systems by up to 50%. Funded by $1.83 million from the Defense Advanced Research Projects Agency (DARPA), the study involves a partnership with the University of Illinois Urbana-Champaign and BAE Systems.
While data center energy consumption is projected to more than double from approximately 420 terawatt-hours in 2024 to over by 2030, artificial intelligence plays a major role in this growth. Researchers note that up to 30% of electrical energy is wasted during artificial intelligence model training due to software and hardware mismatches.
Instead of testing thousands of hardware and software configurations, the developed system will enable developers to estimate energy usage across all stages, from data processing to model training and real-world production. In the project, precise energy measurement tools combining physical sensors with software analysis will be built and published as open-source documents and software.
Sources
- news.engin.umich.edu · Link