Tensor operations involve math on multi-dimensional data structures, like arrays that can represent images or text in artificial intelligence (AI). These operations form the core of tasks such as recognizing pictures or understanding language. Traditional computers, including graphics processing units or GPUs, which are specialized chips for handling large data, process these step by step. This method uses a lot of energy and slows down as data grows bigger. A new approach changes this by using light to do the work all at once.
Researchers led by Aalto University have encoded data into light waves by changing their amplitude and phase. When these waves mix, they naturally multiply matrices and tensors, which are key steps in deep learning algorithms that train AI systems. By adding different colors of light, or wavelengths, the method handles even more complex tensors. This happens as light travels, without needing electronic switches or active controls, making it simple and passive.
How light enables parallel computing
Think of it like checking packages at customs: instead of one at a time, all checks and sorting occur together in one go. Light acts as hooks that link inputs to outputs instantly. This single-shot computing runs at light speed, far faster than electronic methods.
The technique works on various optical setups and could integrate into photonic chips, which use light for processing instead of electricity. This promises low-power systems for tough AI jobs. Experts predict it could reach company hardware in three to five years, boosting fields like image analysis and language tools.
"Our method performs the same kinds of operations that today’s GPUs handle, like convolutions and attention layers, but does them all at the speed of light," note the researchers in a press release. "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 researchers have described the methods and results of this study in a paper published in Nature Photonics.