Photo/Illutration A green laser rewrites a semiconductor’s functions using an NTT Research Inc. team’s “programmable nonlinear photonics” technology. (Provided by NTT Research Inc.)

A research team from NTT Research Inc. has developed technology that is expected to help reduce the massive electricity consumption of data centers handling artificial intelligence.

The researchers, divided across Silicon Valley in California, where NTT Research is based, Cornell University and other institutions turned to light instead of electricity for greater control when assigning functions to the substrate of a microchip known as a photonic integrated circuit. 

Their findings have been published in the British scientific journal Nature.

Advancements in semiconductor chips mean circuit lines are now so small that they are measured on a nano scale—about one hundred-thousandth the thickness of a human hair.

The problem is that the more these circuits shrink, the more heat they generate, requiring large amounts of electricity to keep them cool. The rapid increase in data centers’ power consumption due to the spread of AI has become a global challenge.

POTENTIAL FOR MACHINE LEARNING

The research team tackled this issue by pursuing integrated circuits technology that uses light instead of electrons; another bonus is that light can perform computations faster and more efficiently than electronic circuits.

This latest research achievement is being called “programmable nonlinear photonics.” It makes it possible to incorporate multiple functions on a single substrate and switch the functions as needed.

The technology can also be applied to machine learning, including large language models (LLMs), which are core AI technologies.

Ryotatsu Yanagimoto, a researcher at NTT Research, said, “Computations that computers currently perform using large amounts of electricity can be done more efficiently with light.”

Because semiconductors are complex and highly precise, defects and malfunctions frequently occur during the manufacturing process. Increasing the proportion of fully functional products is tied to the success or failure of a maker.

Previous photonic integrated circuit chips faced a similar quality challenge since circuits could not be modified once burned in. However, the new findings are also seen as applicable to this problem.

The challenge for practical use in AI is whether photonic integrated circuits can be scaled up.

“It depends on the cost, but physically it is possible,” Yanagimoto said.

Related News