A latest AI chip powered by light, not electricity, can provide massive energy savings even as matching conventional performance.
A team of engineers has generated a breakthrough computer chip that depend on light in place of electricity to carry out one of the most energy-traumatic functions in artificial intelligence: figuring out images and identifying patterns.
By shifting to light, the chip needs far less energy, obtaining efficiencies which are 10 to 100 times big than these latest chips running the identical type of calculations. This development should relaxation the huge strain AI places on power grids while also helping the development of more superior and capable AI models.
The main operation included is referred to as “convolution,” a procedure central to how AI explains images, videos, and even written language. At present, convolution is highly resource-intensive and time-taking. The latest design solves this through blending lasers and tiny lenses directly onto circuit boards, permitting the chip to finish these computations with considerably less energy and greater speed.
In early trials, the chip accomplished about 98% accuracy whilst spotting handwritten digits, matching the overall performance of conventional electronic chips.
A Leap Forward in AI Efficiency
“Accomplishing a main machine learning computation at near-zero energy is a leap ahead for future AI systems,” stated study leader Volker J. Sorger, Ph.D., the Rhines Endowed Professor in Semiconductor Photonics on the University of Florida. “This is vital to maintain scaling up AI capabilities in future years to come.”
“This is the first time anyone has placed this kind of optical computation on a chip and carried out it to an AI neural network,” stated Hangbo Yang, Ph.D., a research associate professor in Sorger’s group at UF and co-author of the reseach.
Sorger’s group teamed up with researchers at UF’s Florida Semiconductor Institute, the University of California, Los Angeles and George Washington University on study. The team published their findings on September 8, in the journal Advanced Photonics.
The prototype chip uses 2 units of miniature Fresnel lenses utilizing standard production techniques. These 2-dimensional versions of the identical lenses found in lighthouses are just a fraction of the width of a human hair. Machine learning data, including from an image or different pattern-recognition tasks, are transform into laser light on-chip and exceeded via the lenses. The results are then transformed back into a digital signal to finish the AI task.
Advantages of Optical Computing
This lens-primarily based convolution system is not simplest more computationally efficient, but it also decrease the computing time. Using light in place of electricity has different advantages, too. Sorger’s group designed a chip that could use different colored lasers to process multiple data streams in parallel.
“We can have multiple wavelengths, or colors, of light passing via the lens at the same time,” Yang stated. “That’s a main benefit of photonics.”
Chip producers, which include industry chief NVIDIA, already include optical elements into other parts of their AI systems, that can make the addition of convolution lenses more seamless.
“In the close to future, chip-based optics turns into a main part of every AI chip we use each day,” stated Sorger, who’s also deputy director for strategic projects on the Florida Semiconductor Institute. “And optical AI computing is next.”
