Researchers Develop a New Approach for Energy-Efficient AI "at the Speed of Light"

An international team of researchers, led by Yufeng Zhang from Aalto University's Photonics Group, has claimed a breakthrough in improving the performance and energy efficiency of artificial intelligence — unlocking a new approach to tensor operations that operates "at the speed of light itself."

"Our method performs the same kinds of operations that today's GPUs [Graphics Processing Units] handle, like convolutions and attention layers, but does them all at the speed of light," Zhang says of his team's research. "Instead of relying on electronic circuits, we use the physical properties of light to perform many computations simultaneously."

Current artificial intelligence approaches are near-exclusively based on highly-parallel computation using processors that were originally developed for rendering 3D graphics. Even dedicated neural processing units (NPUs), designed specifically for neural network operations, are closer to a traditional graphics chip than the human brain — putting limits on their performance and efficiency.

This is where the team's research comes into play. Swapping electronics for photonics — where signals are processed as light, rather than electrical impulses — isn't a new concept, but Zhang and colleagues say they have found a way to encode data into the amplitude and phase of light waves in a way that allows them to interact and combine to directly perform matrix and tensor multiplications.

"Imagine you’re a customs officer who must inspect every parcel through multiple machines with different functions and then sort them into the right bins," Zhang explains of how the approach improves efficiency. "Normally, you'd process each parcel one by one. Our optical computing method merges all parcels and all machines together — we create multiple 'optical hooks' that connect each input to its correct output. With just one operation, one pass of light, all inspections and sorting happen instantly and in parallel."

"This approach can be implemented on almost any optical platform," adds Zhipei Sun, leader of Aalto University's Photonics Group, of the team's research. "In the future, we plan to integrate this computational framework directly onto photonic chips, enabling light-based processors to perform complex AI tasks with extremely low power consumption."

The team's work has been published in the journal Nature Photonics; Zhang says that the technology could be deployed on existing or specifically-designed hardware within the next three to five years.