Light-based chips could help quench AI’s growing thirst for energy


“What we have is incredibly simple,” said Tianwei Wu, lead author of the study. “We can reprogram it, change the laser pattern instantly.” The researchers used this system to design a neural network that successfully discriminates vowel sounds. Most photonic systems need to be trained before they can be built, since training essentially involves reconfiguring connections. But since this system can be easily reconfigured, the researchers trained the model after it was installed on the semiconductor. They now plan to increase the size of the chip and encode more information in different colors of light, which should increase the amount of data it can handle.

This progress is such that even Pasaltis, who created a facial recognition system in the 90s, considers it impressive. “Our craziest dreams 40 years ago were very modest compared to what actually happened.”

The first rays of light

Although optical computing has advanced rapidly in the past several years, it is still far from displacing the electronic chips that run neural networks outside of laboratories. Research papers announce photonic systems that perform better than electronic systems, but they typically run smaller models using older network designs and smaller workloads. And many of the statistics touted about photonic supremacy don’t tell the whole story, said Bhavin Shastri of Queen’s University in Ontario. “It’s very hard to make an apples-to-apples comparison with electronics,” he said. “For example, when they use lasers, they don’t really talk about the energy to power the lasers.”

Lab systems need to scale up before they can show a competitive advantage. “How big do you have to make it to win?” asked McMahon. The answer: extraordinarily big. That’s why no one can match the chips made by Nvidia, whose chips power many of the most advanced AI systems today. There’s a huge list of engineering puzzles along the way — issues that the electronics side has spent decades solving. “Electronics is starting out with a huge advantage,” McMahon said.

Some researchers believe ONN-based AI systems will first find success in specialized applications, where they offer unique advantages. One promising use is to counteract interference between different wireless transmissions, such as 5G cellular towers and radar altimeters that help planes navigate, Shastri said. Earlier this year, Shastri and several colleagues created an ONN that can sort through various transmissions and pick out the signal of interest in real time and with a processing delay of less than 15 picoseconds (15 trillionths of a second) — less than one-thousandth of the time an electronic system takes, while using less than 1/70 the power.

But McMahon said the grand vision — an optical neural network that could surpass electronic systems for general use — is still worth trying. Last year his group ran simulations showing that within a decade, a large enough optical system could make some AI models 1,000 times more efficient than future electronic systems. “A lot of companies are working hard now to get to a 1.5-fold advantage. A thousand-fold advantage, that would be amazing,” he said. “It’s probably a 10-year project — if it succeeds.”

Origin Story Reprinted with permission Quanta Magazine, Editorially independent publications Simons Foundation Whose mission is to enhance the public’s understanding of science by covering research developments and trends in mathematics and the physical and life sciences.


Please enter your comment!
Please enter your name here

Share post:




More like this

Paris-based VC Breega achieves first close of $75m Africa fund to support pre-seed and seed startups

Paris-based VC firm Breega has seen Africa’s tech ecosystem...

Today’s Wordle hints, answers and help June 20, #1097

Today's Wordle answer shouldn't be too difficult, but if...

Europe is struggling for relevance in the age of AI

This concentration of power is inconvenient for European governments....