Artificial intelligence (AI) systems are becoming the center of technology that enhances everything from facial recognition to language translation. However, as AI models become more complex, they consume a huge amount of electricity and present the challenges of energy efficiency and sustainability. The new chip, developed by researchers at the University of Florida, helps to address this issue using light rather than just electricity to perform one of the most power-hungry tasks of AI. Their research has been reported in Advanced Photonics.
The chip is designed to perform convolution operations, a core machine learning function that allows AI systems to detect patterns in images, videos and text. These operations usually require critical computing power. By integrating optical components directly into the silicon chip, the researchers created a system that uses laser light and microscope lenses to perform convolution.
“The University of Florida has a great opportunity to develop a new line of expertise in the world,” said Volker J. Sauger, Lines Award Professor of Semiconductor Photonics at the University of Florida. “This is important to continue to expand AI capabilities over the next few years.”
In tests, the prototype chips classified handwritten numbers with approximately 98% accuracy, comparable to traditional electronic chips. The system uses two sets of miniature Fresnel lenses, a flat, ultra-thin version of the lighthouse lens, built using standard semiconductor manufacturing techniques. These lenses are narrower than human hair and are etched directly into the tip.
To perform convolution, the machine learning data is first converted to laser light on the chip. Light passes through a Fresnel lens that performs a mathematical transformation. The results are returned to the digital signal to complete the AI task.
“This is the first time that this type of optical calculation has been placed on a chip and applied to AI neural networks,” said Hangbo Yang, an associate professor of research in the UF's Sorger group and co-author of the study.
The team also demonstrated that using lasers of different colors (a technique known as wavelength multiplexing) allows chips to process multiple data streams simultaneously. “Light of multiple wavelengths or colors can pass through the lens at the same time, and have colors,” Yang said. “That's an important advantage of photonics.”
This study was conducted in collaboration with the Florida Semiconductor Research Institute, UCLA, and George Washington University. Sorger said that chip makers such as Nvidia already use optical elements in some parts of their AI systems, which could make integration easier for this new technology.
“In the near future, chip-based optics will become an important part of all AI chips we use every day,” Sauger said. “And then there's optical AI computing.”
