University of Florida researchers created a silicon photonic chip that uses light to perform AI computations 100x more efficiently. The breakthrough reduces energy use, boosts speed, and could transform future AI hardware with near-perfect accuracy.
Artificial intelligence (AI) is rapidly advancing, but its growing energy demands are becoming a major challenge. Training and running large AI models consumes enormous amounts of electricity, raising concerns about cost, sustainability, and scalability. Now, researchers at the University of Florida have unveiled a breakthrough silicon chip that uses light instead of electricity for one of AI’s most power-hungry tasks—potentially making AI up to 100 times more energy efficient.
The new chip, described in Advanced Photonics, is designed to perform convolution operations, a fundamental part of machine learning that allows AI to detect patterns in text, images, and video. These operations typically require heavy computational power, but the Florida team integrated microscopic optical components directly onto a silicon chip, enabling them to carry out these calculations using laser light. The findings on the light-powered AI chip were published in Advanced Photonics, a peer-reviewed journal from SPIE (the International Society for Optics and Photonics).
How it Works
The chip relies on ultrathin Fresnel lenses—miniature versions of lighthouse lenses—that are etched directly into the silicon and are smaller than a human hair. When machine learning data is converted into laser light, it passes through the lenses, which perform the convolution. The output is then converted back into an electrical signal, completing the AI computation with a fraction of the energy normally required.
In tests, the prototype achieved 98% accuracy in classifying handwritten digits—matching traditional electronic chips while consuming drastically less power.
Speed and Efficiency at Scale
Beyond energy savings, the chip can handle multiple data streams at once using wavelength multiplexing, a method where lasers of different colors carry different information simultaneously. “We can have multiple wavelengths, or colors, of light passing through the lens at the same time,” explained co-author Hangbo Yang, a research associate professor at UF. “That’s a key advantage of photonics.”
Future of Optical AI Chips
According to lead researcher Volker J. Sorger, Rhines Endowed Professor in Semiconductor Photonics at UF, this innovation is “a leap forward for future AI systems” and could help AI continue scaling without overwhelming energy demands.
The work was carried out with collaborators from the Florida Semiconductor Institute, UCLA, and George Washington University. With major chipmakers like NVIDIA already using limited optical elements in AI hardware, the UF team believes integration could happen sooner than expected.
“In the near future, optics will be a key component of every AI chip,” Sorger said. “And fully optical AI computing is next.”