MIT Lincoln Lab Tackles AI's Massive Energy Use Through New Tools and Partnerships

New tools can help reduce the massive energy consumption of AI models, without greatly impacting performance. Techniques include capping GPU power and stopping underperforming models early in training.

The MIT Lincoln Lab is pioneering energy-aware computing research and promoting transparency, publishing frameworks for analyzing data centers' carbon footprints.

Incentive misalignment has slowed adoption of green computing, as the AI field races to build bigger models. But conferences now require ethics statements considering climate impacts.

The Lab is partnering with hardware makers like Intel to standardize energy reporting for users. They aim to give developers control over energy use.

Energy savings could allow delayed hardware purchases, cutting embodied emissions. The Lab is helping facilities like Air Force data centers implement these techniques.