New tools are available to help reduce the energy that AI models devour

The MIT Lincoln Laboratory Supercomputing Center (LLSC) is developing techniques to reduce energy use in data centers, particularly for AI models. They have found that power-capping hardware and early stopping during AI training can significantly decrease energy consumption without impacting model performance. The LLSC has also created a software that allows data center owners to set power limits on GPUs. They have developed a model for hyperparameter optimization that reduces energy waste, and an optimizer that matches models with the most carbon-efficient mix of hardware for inference. These interventions have the potential to advance the way AI models are trained and decrease energy consumption by 10-20%. The LLSC team is promoting transparency and sustainability in the industry and is collaborating with manufacturers and the U.S. Air Force to implement their energy-saving techniques in other data centers.