A new optical computing method that performs complex tensor operations in a single flash of light could redefine AI hardware, delivering massive speed gains with a fraction of the energy of today’s electronic systems.
Research: Direct tensor processing with coherent light. Image Credit: Shutterstock AI
Tensor operations underpin nearly every modern computational system, especially artificial intelligence. These operations go far beyond basic arithmetic; they resemble rotating or slicing a Rubik’s cube across multiple dimensions. Classical computers and humans must compute these steps sequentially, but light can perform them simultaneously.
The Limits of Conventional Hardware
As AI models grow in size and complexity, GPUs and other digital processors are hitting bottlenecks in speed, scalability, and energy use. Every major AI task - from vision to language processing - relies on dense, multi-dimensional tensor operations that strain electronic hardware.
Single-Shot Tensor Computing at the Speed of Light
An international team led by Dr. Yufeng Zhang of Aalto University’s Photonics Group has introduced a breakthrough optical method that carries out complex tensor computations in a single propagation of light. This “single-shot” approach performs operations analogous to convolutions and attention layers used in deep learning, but entirely through the physics of light. The research was published in Nature Photonics on November 14th, 2025.
By encoding digital data into the amplitude and phase of light waves, the researchers effectively transform numerical information into optical fields. When these fields combine, they inherently execute mathematical operations such as matrix and tensor multiplications. Using multiple wavelengths extends the method to higher-order tensor processing.
Instant Parallelism Through Optical Interactions
Dr. Zhang compares the system to merging all parcels and inspection machines in a customs facility so that everything is processed and sorted in one instant. The optical system creates many “hooks” between inputs and outputs, completing all tasks in parallel during a single pass of light.
Because the computation happens passively as light travels, the method requires no electronic switching or active control, significantly simplifying system design.
Toward Photonic Chips for AI
Professor Zhipei Sun, leader of the Photonics Group at Aalto University, notes that the approach is compatible with many optical platforms and is intended for integration into photonic chips. Such processors could deliver complex AI computation with extremely low energy consumption.
Dr. Zhang expects integration into commercial hardware within 3–5 years, creating a new generation of optical computing systems capable of dramatically accelerating AI workloads across diverse applications.
Source:
Journal reference:
- Zhang, Y., Liu, X., Yang, C., Xiang, J., Yan, H., Fu, T., Wang, K., Su, Y., Sun, Z., & Guo, X. (2025). Direct tensor processing with coherent light. Nature Photonics, 1-7. DOI: 10.1038/s41566-025-01799-7, https://www.nature.com/articles/s41566-025-01799-7
AI Meets Metal: Rowan’s DEHub Redefines Smart Manufacturing With Supercharged 3D Printing