A new generation of computer chips that drive light rather than electricity could revolutionize artificial intelligence, researchers say.
A team of US engineers have developed prototype devices that can perform AI calculations that are 10-100 times more energy efficient than today's best chips.
Breakthroughs concentrate on convolution, one of the most power-hungry operations in machine learning. This process allows AI systems to recognize patterns in photos, videos, and even written text, but consume a lot of power with traditional processors.
read more: From brain bandages to reborn teeth – and Google's $425 million headache
Instead, the new design uses a microscope lens etched directly onto the laser and circuit board. In clinical testing, the chip was consistent with the accuracy of the electronic chip, achieving 98% success when identifying handwritten numbers while using only a portion of the energy.
Hangbo Yang, an associate professor of research at the University of Florida and research co-author, said:
Volker J. Sorger, a lead researcher at the University of Florida State Semiconductor Research Institute, described the advance as “a leap for AI systems of the future.” He added:
“To continue to expand AI capabilities over the next few years, it is important to perform important machine learning calculations with energy near zero.”
The study, published in Advanced Photonics on September 8th, included a collaboration between the University of Florida, UCLA and George Washington University.
Read again: AI avatars bring comfort, but cannot replace sadness
How the chip works
This prototype integrates two sets of ultra-thin Fresnel lenses. This is the largest version of what you can see in the lighthouse, but it's just a small portion of human hair. Machine learning data is first converted to a laser light-on-chip, then passed through the lens, then converted to a digital signal to complete the task.
The use of light has another advantage. Lasers of different colors can be used simultaneously to process multiple data streams in parallel, a feature known as wavelength multiplexing.
“It can have multiple wavelengths or colors of light passing through the lens at the same time. That's an important advantage of photonics,” explained Yang.
Impact on the industry
Chipmakers, including market leader Nvidia, are already using optical components in some of their AI hardware. Sauger believes that convolutional lenses will soon be able to be integrated into mainstream products.
“In the near future, chip-based optics will become an important part of all AI chips that we use every day,” he said. “And then there's optical AI computing.”
