The MIT Lincoln Laboratory Supercomputing Center (LLSC) is developing techniques to reduce energy consumption in data centers, particularly in relation to artificial intelligence (AI) models. They have found that simple changes like power-capping hardware can reduce energy consumption by 12-15% without significantly impacting model performance. The LLSC has also developed software that allows data center owners to set power limits across their systems. Additionally, they have created a comprehensive framework for analyzing the carbon footprint of high-performance computing systems. The team is also focusing on making AI-model development more efficient by using early stopping techniques, which have led to an 80% reduction in energy used for model training. They have also developed an optimizer that matches AI models with the most carbon-efficient mix of hardware for inference. The energy saved by implementing these interventions reduces costs and aligns with the growing awareness of green computing. The team is pushing for transparency and reporting tools to show AI developers their energy consumption and is working with hardware manufacturers to standardize data collection. They aim to help other data centers apply these interventions and provide energy-aware options to AI developers.
