Could Computing With Light Finally Make AI Profitable?

A paper published this week in the journal Science Advances demonstrates an optical computing concept that could greatly accelerate AI, while reducing energy costs and heat production. Intriguingly, it does all this by way of an “infinity mirror.”
Optical computers have several intrinsic advantages, provided they work. Not only does light move very quickly, reducing latency, but it can also encode information in multiple ways. By encoding information in light’s wavelength, phase, polarization state, and more, an optical computer can greatly increase data bandwidth compared to pure electronic systems.
Light can even be split and recombined in ways that mimic complex mathematical operations, physically dividing, transforming, and summing waves as a stand-in for doing the same operations on abstract numbers. Light has inherent parallelism, meaning it can be split and recombined to perform precisely the matrix multiplications needed for neural networks.
However, to organize these math operations so they can do useful work, let alone to structure them so they work like a neural network diagram, scientists need to be able to perform nonlinear operations. This basically means there needs to be an intensity-dependent aspect to the treatment of light, enabling all the operations of a computer. That includes all-important logic gates such as AND, OR, and NOT, as well as the functions applied by the nodes of a neural network.
The techniques that enable such nonlinear diffraction and light control have historically been expensive, power-hungry, or both, offsetting the reasons for building an optical network in the first place.
That’s where this team’s infinity mirror comes in. Very simply, the team has sandwiched a tiny transparent LCD display between two partial mirrors that reflect only certain polarizations of light. By inputting this polarization of light, they can trap it inside the sandwich and cause it to be transmitted through the LCD over and over again.
Researchers can get the LCD display to modulate the amplitude of the light that passes through it, so some of the affected light will not be reflected and will escape the sandwich. This allows the signal to get clearer and clearer with each pass through the LCD, and enables the sort of intensity-dependent, nonlinear response needed.
The team seems to think that early applications for their technology could be ready to roll out within as little as a few years, likely in simple chips made for sensing in productive environments. We’ll see if the concept scales to the point of fulfilling more complex consumer-grade AI needs.
© 2001-2026 Ziff Davis, LLC., a Ziff Davis company. All Rights Reserved.
ExtremeTech is a federally registered trademark of Ziff Davis, LLC and may not be used by third parties without explicit permission. The display of third-party trademarks and trade names on this site does not necessarily indicate any affiliation or the endorsement of ExtremeTech. If you click an affiliate link and buy a product or service, we may be paid a fee by that merchant.

source