This article digs into a fresh breakthrough in optical computing. Researchers now enable true parallel matrix–matrix and tensor–matrix multiplications using free-space optics.
Published in Nature Photonics in January 2026, this work tackles a bottleneck that’s kept optical systems stuck with simpler linear algebra tasks. By breaking through that wall, the research nudges optical hardware closer to being practical and energy-efficient for machine learning and scientific computing.
Why Matrix–Matrix Multiplication Matters in Optical Computing
Linear algebra operations form the backbone of today’s computing workloads—think AI or physics simulations. Optical computing has long held out the promise of being faster and more energy-efficient than electronics, but it hasn’t quite delivered on that hype because it struggled with certain calculations.
Most optical architectures up to now could only handle vector–matrix multiplication. If you wanted matrix–matrix multiplication? You had to rely on indirect workarounds, which kind of defeated the point of using optics in the first place.
The Bottleneck of Multiplexing-Based Approaches
Old-school optical methods leaned on wavelength-division or time-division multiplexing to mimic matrix–matrix multiplication. Clever, sure, but these tricks came with a bunch of extra baggage.
Systems would run into:
All this made it tough for optical computing to scale up or catch on for big numerical workloads.
A Free-Space Optical Breakthrough
This new method flips the script. Now, parallel matrix–matrix and tensor–matrix multiplications happen in a single optical operation.
Carlos A. RÃos Ocampo and Nathan Youngblood developed this approach. They use free-space optics to twist and steer light so it directly encodes multi-dimensional linear algebra.
They managed to do this without falling back on time or wavelength multiplexing. The whole computation happens in one pass of light through the system.
How Free-Space Optics Enable True Parallelism
Free-space optical systems can shape, modulate, and interfere light across two spatial dimensions. By playing with these properties, the team shows how entire matrices and tensors interact optically all at once.
This marks a real shift—from sequential or multiplexed optical operations to true parallel computation. It brings optical hardware a lot closer to how modern algorithms actually work.
Implications for Machine Learning and Scientific Computing
Efficient large-scale linear algebra is at the heart of tons of computational fields. The authors point out that their method could boost both throughput and energy efficiency in a big way.
Potential upsides include:
These benefits matter a lot for optical neural networks, which the paper highlights as a fast-moving area.
Positioning Within the Broader Photonics Landscape
The article puts this advance next to recent progress in photonic accelerators and optical neural networks. Earlier systems showed solid performance for certain tasks, but they still hit a wall with the vector–matrix paradigm.
By breaking what the authors call a fundamental bottleneck, this new approach opens up more problems for optical hardware to tackle.
Transparency, Patents, and Future Outlook
The authors openly share their competing interests related to existing patents in optical computing. It’s a standard move for transparency, but it also hints at the technology’s commercial potential.
This work appeared on 6 January 2026 in Nature Photonics. It points to a real step forward for high-performance, scalable optical computing systems.
There’s still quite a bit of engineering and validation ahead. Still, the team’s demonstration of parallel matrix–matrix multiplication in free space stands out as a milestone that could push optical computing beyond the lab and into real-world scientific and industrial use.
Here is the source article for this story: Multiplying matrices in a single pass with light