A Single Flash of Light Could Transform Machine Intelligence

A customs officer sorting parcels usually checks each one individually, running it through various machines before placing it in the right bin. But what if you could inspect all the parcels simultaneously, merging every machine and sorting operation into a single instant? That’s essentially what researchers at Aalto University have accomplished with artificial intelligence calculations, swapping electronic circuits for beams of light.

In work published November 14 in Nature Photonics, an international team led by Dr. Yufeng Zhang demonstrated that the mathematical operations powering modern AI can be performed in a single pass of coherent light. The approach, called POMMM (parallel optical matrix-matrix multiplication), mimics what GPU chips do today but uses the physical properties of light waves instead of electricity flowing through silicon.

The implications could be significant. Every AI system, from facial recognition to ChatGPT, relies on massive numbers of tensor operations, a type of multidimensional arithmetic that forms the computational backbone of neural networks. As AI models grow larger and more complex, conventional processors are hitting practical limits on speed and energy consumption. Light-based computing offers a potential way forward.

Encoding Numbers Into Waves

The core insight behind POMMM involves transforming abstract mathematics into physical reality. In standard matrix multiplication, you take rows from one matrix and columns from another, multiply corresponding elements together, then add up the results. GPUs handle this by cycling through operations sequentially, even with parallel processing.

Zhang’s team found a way to make light perform all those operations at once. They encoded the numbers from one matrix into the amplitude and phase of a light field, essentially turning data into the brightness and timing of light waves. Each row of numbers gets tagged with a distinct wave pattern. When this light passes through a carefully designed sequence of lenses, the physics of wave propagation automatically performs the necessary multiplications and additions.

“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,” says Dr. Zhang. “Instead of relying on electronic circuits, we use the physical properties of light to perform many computations simultaneously.”

The setup uses spatial light modulators, devices that can control light’s properties across a two-dimensional field, along with cylindrical lenses that shape how the light propagates. One modulator encodes the first matrix, applying different wave patterns to each row. The lens arrangement then superposes all the rows in a way that can be untangled later. A second modulator encodes the transpose of the second matrix, and a final set of lenses separates out the results. A high-resolution camera captures the output, with each bright spot corresponding to one element of the final result matrix.

From Lab Prototype to Neural Networks

Because POMMM produces standard matrix multiplication results, the researchers could test it as a drop-in replacement for GPU calculations. They trained convolutional neural networks and vision transformer models on conventional hardware, then ran the same architectures through both simulated and physical optical systems.

The consistency was striking. Across various matrix sizes and datasets like MNIST handwritten digits and Fashion-MNIST clothing images, the optical results closely matched digital baselines. Mean absolute errors stayed below 0.15, and normalized root mean square errors remained under 0.1 for large samples of random matrices.

“Imagine you’re a customs officer who must inspect every parcel through multiple machines with different functions and then sort them into the right bins,” Zhang explains. “Normally, you’d process each parcel one by one. Our optical computing method merges all parcels and all machines together, we create multiple ‘optical hooks’ that connect each input to its correct output. With just one operation, one pass of light, all inspections and sorting happen instantly and in parallel.”

The team also explored robustness. When optical imperfections introduced larger errors, they trained models directly using the POMMM framework rather than trying to perfectly replicate GPU behavior. High-quality inference remained possible. In some cases where nonlinear activation functions were effectively disabled, POMMM-trained models actually outperformed their GPU counterparts, suggesting the optical approach might offer unexpected advantages for certain architectures.

Complex Numbers and Future Chips

The researchers extended their framework to handle complex-valued matrices by using multiple wavelengths of light. They encoded real and imaginary parts onto different colors, demonstrating complex matrix multiplication as a tensor operation in a single propagation. This multi-wavelength approach opens pathways to even higher-order tensor processing.

The current prototype, assembled from off-the-shelf optical components, achieves modest practical energy efficiency at about 2.62 billion operations per joule. That’s not yet competitive with optimized electronic hardware. However, the underlying operations are passive optical transformations, requiring no active control once the light is propagating. This suggests substantial efficiency gains could come from integration into dedicated photonic chips.

Professor Zhipei Sun, who leads Aalto University’s Photonics Group, emphasizes the platform’s flexibility. “This approach can be implemented on almost any optical platform,” he says. “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.”

Zhang and colleagues estimate their approach could be integrated into commercial optical computing platforms within three to five years. Major technology companies are already developing photonic hardware, and POMMM’s compatibility with standard GPU architectures means neural networks designed for conventional chips could theoretically run on optical processors without requiring custom redesign.

The work arrives at a moment when AI’s appetite for computing power continues to accelerate. Training large language models can consume megawatt-hours of electricity. Inference at scale presents similar challenges. If optical computing can deliver on its promise of massive parallelism with lower energy consumption, the approach demonstrated by Zhang’s team could provide a foundation for the next generation of AI accelerators.

For now, the research represents a proof of concept showing that direct tensor processing with coherent light is feasible. The path from laboratory demonstration to commercial deployment involves considerable engineering, but the fundamental physics appear sound. Whether AI at the speed of light becomes practical reality or remains an elegant curiosity will depend on how well photonic integration can match the theoretical advantages the Aalto team has demonstrated.