Light-based Chip Boosts AI Power Efficiency Up to 100-Fold

A University of Florida team has created a new chip that uses light alongside electricity to handle power-intensive AI tasks like image recognition, achieving 10 to 100 times better efficiency than current chips. This could reduce the strain on power grids from AI's growing electricity demand while enhancing model performance, reports Techxplore. 

AI systems, vital for facial recognition and translation, consume vast energy as they grow complex, challenging sustainability. The new chip tackles this by using light for convolution operations—key to pattern detection in images and text—integrating optical components on a silicon chip with laser light and microscopic lenses to cut energy use and boost speed.

Volker J. Sorger, the study leader, called it a "leap forward" for energy-efficient AI, noting the prototype achieved 98% accuracy in digit classification using miniature Fresnel lenses etched onto the chip. The process converts data into light, performs transformations via lenses, and converts results back to digital signals.

Co-author Hangbo Yang highlighted the chip's ability to process multiple data streams simultaneously using different laser colors (wavelength multiplexing), a key photonic advantage. Collaborating with the Florida Semiconductor Institute, UCLA, and George Washington University, the team sees this as a step toward integrating optical AI into everyday chips, with manufacturers like NVIDIA potentially adopting it soon.