An international research team led by Yufeng Zhang from Aalto University’s Photonics Group has unlocked a new approach to tensor operations that operates “at the speed of light itself,” claiming a breakthrough in improving the performance and energy efficiency of artificial intelligence.
“Our method performs the same kinds of operations as today’s GPUs. [Graphics Processing Units] “We do things like convolution and attention layers, all at the speed of light. Instead of relying on electronic circuits, we use the physical properties of light to perform many calculations simultaneously,” Zhang says of his team’s work.
An international team claims a breakthrough in AI efficiency by processing tensor operations directly as light. (📷: Photonics Group/Aalto University)
Current artificial intelligence approaches are based almost exclusively on highly parallel computation using processors originally developed for rendering 3D graphics. Even dedicated neural processing units (NPUs) designed specifically for neural network operations have performance and efficiency limitations because they are more similar to traditional graphics chips than the human brain.
This is where team research comes in handy. Replacing electronics with photonics (where signals are processed as light rather than electrical impulses) is not a new concept, but Zhang and his colleagues say they have discovered a way to encode data into the amplitude and phase of light waves, which can then be interacted and combined to directly perform matrix and tensor multiplications.
“Imagine you are a customs officer and you have to use multiple machines with different functions to inspect every parcel and sort them into the appropriate boxes,” Zhang explains how this approach increases efficiency. “Typically, we process each package one at a time. Our optical computing approach combines every package and every machine. We create multiple ‘optical hooks’ that connect each input to the correct output. In just one operation, one pass of light, all the inspection and sorting happens instantly and in parallel.”
The team’s research is claimed to be significantly more energy efficient than GPU-based devices from companies such as NVIDIA and AMD. (📷: NVIDIA)
“This approach can be implemented on almost any optical platform,” added Zhipei Sun, leader of Aalto University’s photonics group, who worked on the team’s research. “In the future, we plan to integrate this computational framework directly into photonic chips, allowing light-based processors to perform complex AI tasks with very low power consumption.”
The team’s results were published in a magazine natural photonics; Zhang said the technology could be deployed on existing or specially designed hardware within the next three to five years.
