Want to call someone smart? These days, we often compare them to a computer. But originally, the term “computer” referred to mathematicians—like those in the film “Hidden Figures”—whose calculations were crucial to the success of NASA’s first missions.

Modern electronic computers, despite their speed, still lag behind optical computing, where data is processed not by electrons but by light. This allows for speeds approaching that of light itself.

Why Faster Computing Matters

The need for faster calculations isn’t just about users wanting to watch movies, video chat, play VR games, and 3D print all at once. The main driver is the explosive growth of data in our digital world. Standard graphics processors (GPUs) in regular computers can’t keep up with such massive information flows—they can’t scale or operate fast enough. Additionally, as noted by The Smithsonian Magazine and Sustainability Magazine, AI data centers packed with GPUs consume enormous amounts of electricity (mostly from non-renewable sources) and water, often in drought-prone regions.

A Breakthrough in Optical Computing

Researchers Yufeng Zhang (Photonics Group, Aalto University) and Xiaobin Liu (Chinese Academy of Sciences, Changchun) have introduced a new computational method in Nature Photonics. Their approach uses a single pass of light—so-called “single-pass tensor computations at the speed of light in parallel optical matrix multiplication” (POMMM).

This method brings us a step closer to creating artificial general intelligence. As Zhang explains, their technique performs the same operations as modern GPUs (such as convolutions and attention layers), but at the speed of light. Instead of traditional binary coding with electronic circuits using ones and zeros, the researchers use the amplitude and phase of light waves to store, process, and transmit data. This not only saves energy but also dramatically increases bandwidth and processing speed, enabling many processes to run simultaneously.

How Optical Data Processing Works

The interaction of light fields naturally performs mathematical operations, including tensor multiplications. Tensor processing organizes data into multidimensional arrays (tensors), which is fundamental for deep learning algorithms, data analysis, natural language processing, and image recognition.

However, current optical methods struggle with tensor tasks, limiting their use in neural networks and other complex applications. The new method from the Aalto team offers a significant breakthrough: by using multiple light wavelengths, it can handle even the most complex tensor operations.

An Analogy: Sorting Packages

To illustrate the difference between electronic and optical computing, Zhang uses the analogy of sorting packages. Imagine you’re a customs officer who needs to check each package on several machines with different functions, then sort them into the right bins. Normally, you’d process each package one by one. The optical method combines all the packages and machines, creating many “optical hooks” that connect every input to the correct output. In a single operation, with one pass of light, all checks and sorting happen instantly and in parallel.

Parallel Tensor Processing

Parallel tensor processing (tensor model parallel, TP) is a strategy in deep learning where multiple devices handle components of a single model, allowing larger models to be computed faster. The more participants, the easier the work—or in this case, the faster the processing, approaching the speed of light.

Implementation Prospects

According to Jipei Sun, head of the Photonics Group at Aalto University, the new method can work on almost any optical platform. The team plans to integrate the computational structure directly into photonic chips, so that light-based processors can tackle complex AI tasks with minimal energy consumption.

If successful, the method is expected to be implemented for integration with existing equipment and major platforms within five years. This will usher in a new generation of optical computing systems, dramatically accelerating the performance of complex artificial intelligence tasks across various fields.