Light-Based Computing Crushes Matrix Math

The Computer That Runs on Light

Your laptop is a bottleneck. Every time you ask it to process a matrix—and you do, constantly, every time you run a neural network, every time you filter a photo, every time you ask Siri a question—it sends electrons through silicon wires, heating up, slowing down, hitting walls. The matrix is the workhorse of modern computing, and electrons are terrible at it.

But photons? Photons are good at math. They just needed someone to teach them.

In 2022, a team led by Hailong Zhou at Huazhong University of Science and Technology published a review in Light Science & Applications that mapped out exactly how photonic matrix multiplication works, and why it might finally break the bottleneck that has been strangling artificial intelligence, signal processing, and pretty much every computationally hungry task you can name (Zhou et al., 2022). Their paper is not a single breakthrough. It is a roadmap. And the destination is a computer that does its heaviest lifting not with electricity, but with light.

Why Matrices Matter More Than You Think

Matrices are not just math class abstractions. They are the language of modern computation. Every neural network layer is a matrix multiplication. Every image transformation, every audio filter, every optimization in machine learning—all of it comes down to multiplying one grid of numbers by another grid of numbers. The authors put it bluntly: "Matrix computation, as a fundamental building block of information processing in science and technology, contributes most of the computational overheads in modern signal processing and artificial intelligence algorithms" (Zhou et al., 2022).

That overhead is enormous. When you train GPT-4, you are performing trillions of matrix multiplications. Each one costs energy. Each one generates heat. Each one takes time. And the problem is getting worse, not better. Moore's Law is slowing down. Transistors are approaching atomic limits. But our appetite for computation is not slowing down at all.

The Physics of Photonic Multiplication

Here is the core insight: when you send a laser beam through a specially designed optical system, the light itself performs the matrix multiplication. No transistors. No logic gates. No electrons moving through silicon. Just photons interfering with each other, adding and subtracting, doing math at the speed of light.

Zhou and his colleagues describe three main methods for doing this. The first is called plane light conversion. Imagine sending a laser beam through a series of carefully shaped optical elements. The light enters as one pattern and exits as a different pattern. That transformation is mathematically equivalent to multiplying the input vector by a matrix. The authors explain that "the plane light conversion method uses free-space optics to perform matrix multiplication by shaping the wavefront of light" (Zhou et al., 2022). It is elegant. It is fast. And it is already being used in some experimental systems.

The second method uses Mach-Zehnder interferometers. This is a fancy name for a simple idea: split a beam of light into two paths, let them interfere with each other, and measure the result. By chaining many of these interferometers together in a mesh, you can build a programmable optical circuit that performs arbitrary matrix multiplications. The authors note that "the Mach-Zehnder interferometer method uses a mesh of interferometers to implement arbitrary unitary matrix transformations" (Zhou et al., 2022). This is the approach that companies like Lightmatter are betting on for commercial photonic chips.

The third method is wavelength division multiplexing. Instead of splitting light spatially, you split it by color. Each wavelength carries a different piece of data. You modulate them, combine them, and then separate them again—and the whole process performs a matrix multiplication in the frequency domain. The authors describe how "the wavelength division multiplexing method uses multiple wavelengths to encode different data channels and performs matrix multiplication through wavelength-selective components" (Zhou et al., 2022). This approach is particularly promising for telecommunications, where you already have fiber optics carrying multiple wavelengths.

The Speed Advantage Is Not What You Think

Here is where it gets counterintuitive. You might assume that photonic computing is faster because light moves at the speed of light. That is true, but it is also misleading. Electrons in a wire also move at near light speed. The real bottleneck is not the speed of individual operations. It is the energy cost of moving data around.

In a conventional computer, every time you perform a matrix multiplication, you have to fetch data from memory, move it to the processor, do the math, and store the result back. Each of those steps costs energy. And the energy cost of moving data is often larger than the energy cost of computing on it. This is called the von Neumann bottleneck, and it is the single biggest limiter of modern computing performance.

Photonic computing sidesteps this problem entirely. Because the light itself does the math, you do not need to shuttle data back and forth between memory and processor. The computation happens in flight. The authors emphasize that "photonic matrix multiplication has much potential to expand the domain of telecommunication and artificial intelligence benefiting from its superior performance" (Zhou et al., 2022). That "superior performance" comes from eliminating the data movement bottleneck, not from making individual operations faster.

What This Means for Artificial Intelligence

Neural networks are hungry. They are hungry for matrix multiplications, and they are hungry for energy. Training a single large language model can consume as much electricity as a small town. Inference—the act of actually using the model—is cheaper but still expensive. And the trend is toward bigger models, not smaller ones.

Photonic accelerators could change this calculus. Because they perform matrix multiplications with orders of magnitude less energy, they could make it feasible to run large neural networks on devices that currently cannot handle them. Think smartphones doing real-time language translation without a cloud connection. Think autonomous cars processing sensor data with a fraction of the power they currently use. Think medical imaging systems that analyze scans in real time without a server farm.

The authors review recent advances in applying photonic matrix multiplication to artificial neural networks. They describe how "photonic accelerators are designed to accelerate specific categories of computing in the optical domain, especially matrix multiplication, to address the growing demand for computing resources and capacity" (Zhou et al., 2022). The key word is "accelerator." Photonic chips are not meant to replace your laptop's CPU entirely. They are meant to handle the specific, energy-intensive tasks that conventional processors are bad at—and matrix multiplication is at the top of that list.

The Signal Processing Angle

Before artificial intelligence got all the attention, the main application for photonic matrix multiplication was signal processing. Radars, communications systems, and medical imaging all rely on matrix operations to filter, transform, and analyze signals. And these systems have a problem: as data rates increase, conventional electronics cannot keep up.

The authors describe how photonic matrix multiplication is already being used for optical signal processing. They note that "photonic matrix multiplication has much potential to expand the domain of telecommunication" (Zhou et al., 2022). In telecommunications, you are already dealing with light. Fiber optic cables carry data as pulses of laser light. But when that data reaches a router or a switch, it has to be converted to electricity, processed, and then converted back to light. Each conversion costs energy and introduces latency. A photonic processor that can operate directly on the optical signal—without converting to electricity—could eliminate that bottleneck entirely.

This is not science fiction. The authors review experimental systems that perform matrix multiplication directly on optical signals. They describe how these systems can process data at rates that would overwhelm conventional electronics. And they outline the path from laboratory demonstrations to practical devices.

The Milestones So Far

The field of photonic matrix multiplication has a history that goes back further than you might think. Zhou and his colleagues trace the development from early demonstrations of optical computing in the 1980s to the modern integrated photonic circuits that are now being commercialized. They identify several key milestones:

• ▸The first demonstration of programmable photonic circuits using Mach-Zehnder interferometer meshes
• ▸The development of wavelength division multiplexing techniques for parallel optical computation
• ▸The integration of photonic components onto silicon chips, making them compatible with existing manufacturing processes
• ▸The demonstration of photonic neural networks that can perform classification tasks with high accuracy

The authors summarize these milestones in their review, showing how each step built on the previous one. They write that "recent research in photonic matrix multiplication has flourished and may provide opportunities to develop applications that are unachievable at present by conventional electronic processors" (Zhou et al., 2022). That last part is crucial. Photonic computing is not just a faster version of what we already have. It enables things that are currently impossible.

What Photonic Computing Cannot Do Yet

Here is where honesty matters. Photonic matrix multiplication is not a magic bullet. It has limitations, and the authors are upfront about them.

First, photonic systems are currently larger and more expensive than their electronic counterparts. A silicon photonic chip requires specialized fabrication processes. It requires precise alignment of optical components. It requires lasers and modulators and detectors that add cost and complexity. The authors note that "the challenges of photonic matrix multiplication include the integration of light sources, modulators, and detectors on a single chip" (Zhou et al., 2022). This is not a fundamental physics problem. It is an engineering problem. But it is a hard one.

Second, photonic computing is not good at everything. It is excellent at matrix multiplication, which is a linear operation. But many computations require nonlinear operations—activation functions in neural networks, for example. Photonic systems can perform some nonlinear operations, but they are harder to implement optically than linear ones. The authors acknowledge that "the nonlinear activation function in photonic neural networks remains a challenge" (Zhou et al., 2022). Hybrid approaches, where photonic chips handle the linear operations and electronic chips handle the nonlinear ones, are likely to be the near-term solution.

Third, photonic systems suffer from noise and loss. Light leaks. Components are imperfect. And every photon that gets lost is information that disappears. The authors discuss how "the loss in photonic circuits limits the scalability of photonic matrix multiplication" (Zhou et al., 2022). For small matrices, this is manageable. For the massive matrices used in modern AI, it becomes a serious constraint.

The Open Question: How Big Can It Get?

The most interesting question the paper raises is not whether photonic matrix multiplication works. It clearly does. The question is how far it can scale.

Can you build a photonic chip that multiplies matrices with millions of elements? Can you integrate thousands of interferometers on a single chip without the losses becoming prohibitive? Can you manufacture these chips at a cost that competes with silicon electronics?

The authors do not pretend to have definitive answers. They describe the current state of the art and outline the path forward. They point to recent advances in integrated photonics, including the development of low-loss waveguides and efficient modulators, as reasons for optimism. But they also acknowledge that "the challenges of photonic matrix multiplication remain significant" (Zhou et al., 2022).

This is not a weakness of the paper. It is a strength. The authors are not selling a solution. They are describing a technology that is real, that works, and that has enormous potential—but that still has a long way to go before it becomes practical.

What This Actually Means

• ▸Photonic matrix multiplication is not a theoretical curiosity. It is a working technology that has been demonstrated in multiple laboratories and is being commercialized by companies like Lightmatter and Lightelligence. The physics is sound. The engineering is the remaining challenge.

• ▸The advantage of photonic computing is not raw speed. It is energy efficiency. By eliminating the data movement bottleneck, photonic accelerators can perform matrix multiplications using orders of magnitude less energy than electronic processors. For applications where energy is the limiting factor—battery-powered devices, data centers, autonomous systems—this is transformative.

• ▸Neural networks are the killer app. Artificial intelligence is currently constrained by the energy cost of matrix multiplication. Photonic accelerators directly address this constraint. The authors review how photonic neural networks have already been demonstrated, and the trajectory is clear: more capable systems are coming.

• ▸Telecommunications will benefit first. Because fiber optic networks already use light, photonic signal processing can be integrated directly into existing infrastructure. The authors describe how photonic matrix multiplication is being applied to optical signal processing, and this is likely to be the first widespread commercial application.

• ▸The technology is not ready for general-purpose computing. Photonic accelerators are specialized devices. They are excellent at matrix multiplication and poor at almost everything else. The near-term future is hybrid systems where photonic chips handle the heavy lifting of matrix operations and electronic chips handle everything else. This is not a limitation. It is the smart way to deploy a new technology.

• ▸The open question is scalability. Can photonic matrix multiplication be scaled to the size and cost required for mainstream adoption? The authors are cautiously optimistic. The engineering challenges are real, but they are being solved. The next decade will determine whether photonic computing becomes a niche technology or a fundamental part of the computing landscape.

Light does not care about Moore's Law. It does not care about transistor sizes or thermal limits. It just does math, at the speed of light, with almost no energy. The question is whether we are smart enough to let it.