UCLA builds optical AI processor that computes at the speed of light

As AI models balloon in size, data centers are running into a massive power wall. Traditional silicon chips rely on shifting electrons through billions of transistors, which generates serious heat and consumes staggering amounts of electricity.

Optical computing offers an alluring alternative: swap electrons for photons to compute faster with a fraction of the power. On January 18, IEEE Spectrum reported that researchers at UCLA, led by optical engineer Aydogan Ozcan, developed a broadband diffractive optical processor that runs neural network operations at the speed of light.

The findings were originally published on January 9 in the journal Advanced Photonics.

Computing by bending light

A diffractive optical neural network doesn't look anything like a graphics card. Instead, it uses a physical stack of patterned layers, each containing thousands of tiny pixels designed to scatter and bend light. These diffractive features act as artificial neurons.

Instead of snapping an image, converting it into digital bits, and churning through matrix multiplications in memory, this optical setup calculates passively. Light reflecting off an object passes through the 3D-printed layers, diffracts, and naturally focuses onto target output sensors. As Ozcan points out, the entire computing process consumes essentially zero power, aside from the light source itself.

Thousands of matrix operations at once

Earlier versions of diffractive networks were monochromatic, meaning a single color of light could only run one matrix multiplication at a time. UCLA's latest breakthrough introduces wavelength multiplexing to process broad spectrums simultaneously.

In their experiments, the scientists 3D-printed three diffractive layers, each packed with 14,400 diffractive features. They successfully demonstrated the system using two submillimeter-wavelength terahertz channels. More importantly, their numerical simulations indicate that the architecture can scale to roughly 2,000 wavelength channels operating in parallel.

That means thousands of heavy matrix operations can occur simultaneously across different frequencies without interference, all traveling at the speed of light.

Why it matters

The AI race is quickly becoming an energy bottleneck, and traditional semiconductor scaling can't keep up forever. By replacing power-hungry electronic logic gates with passive optical elements, this diffractive approach solves the matrix multiplication problem using fundamental physics rather than brute force.

We think this passive optical architecture is particularly compelling for edge sensing. Autonomous vehicles, medical diagnostic scopes, and remote sensing satellites could classify objects and filter signals instantly at the lens level, slashing the need for heavy onboard GPUs.

That said, it isn't about to replace your Nvidia cluster tomorrow. 3D-printed diffractive layers bake neural weights directly into physical plastic, meaning they can't be reprogrammed on the fly like software. Moving from two experimental terahertz channels to 2,000 channels across visible light will also require immense manufacturing precision. Still, demonstrating massively parallel matrix math at zero computing wattage shows where the future of sensing is headed.