New light-based computer chip revolutionizes AI — processes data millions of times faster

As artificial intelligence grows more powerful, so does its appetite for speed and energy. The quest for faster, smarter systems has driven researchers to an unlikely ally—light itself.

A new study by researchers at Tsinghua University introduces a groundbreaking optical feature extraction engine that could revolutionize AI processing. By processing data using light rather than electricity, the technology accelerates computing exponentially while minimizing latency, a major leap toward real-time, energy-efficient AI.

Traditional processors—whether GPUs or TPUs—rely on the flow of electrons through circuits. Electrons, however, are sluggish and generate heat, causing bottlenecks as data demands grow. Photonics, which communicates and processes information with light, has the ability to demolish those limits. Light travels nearly a million times faster than electrical currents and produces significantly less heat, offering a simple path to quicker, cooler, and cleaner computing.

However, creating photonic devices that match or beat electronics has remained elusive. Bulk, expense, and low modulation rates have restricted most existing optical setups. That is why this new optical engine is such a breakthrough—it combines speed, efficiency, and compactness on one chip.

At the heart of the new system is a semiconductor optical amplifier–based Mach–Zehnder interferometer, or SOA-MZI. The setup allows light to perform the mathematical operations underlying deep learning, known as convolutions. It basically processes the information as it goes along, detecting features like edges, patterns, and motion directly in the light signal—without ever being converted back into electricity.

The device also uses wavelength-division multiplexing, or WDM, a method that splits light into a spectrum of colors, each carrying its own data stream. That enables the chip to conduct many calculations in parallel, significantly enhancing its throughput. In the lab, the engine processed data at speeds of up to 10 gigabits per second per channel with latency of just tens of picoseconds—far faster than any electronic processor could possibly hope to achieve.

This speed is not just about bragging rights. It implies that the system can process signals and images in real time, making it ideal for applications like autonomous vehicles, medical imaging, or high-frequency trading—domains where even milliseconds matter.

Professor Hongwei Chen and his team at Tsinghua University developed a second-generation optical feature extraction engine, or OFE2. What they did was solve one of the most vexing problems in the area of optical computing: how to maintain a stable, coherent light source when operating at high speed. They did it by developing an on-chip module with tunable power splitters and precise delay lines, which delivers stable, parallel optical signals.

This arrangement de-serializes incoming information—essentially splitting an ongoing input into many synchronized light waves—allowing parallel, real-time processing. Once the light is formatted, it goes through a diffraction operator, a microscopic device that performs mathematical operations as light travels through it. The result is an operation akin to matrix-vector multiplication, a fundamental building block of modern AI.

Operating at 12.5 gigahertz, OFE2 can finish a whole feature extraction in slightly over 250 picoseconds. Chen has called this a record for optical computing, pushing the field beyond 10 GHz in real-world applications.

To illustrate the potential of what their device could do, the researchers used OFE2 on several practical applications. In image processing, the engine was utilized for extracting edges and faint details in images and generating two complementary maps resembling relief and engraving effects. These optical features enhanced image classification accuracy and boosted the accuracy of medical image segmentation, such as in finding organs in CT scans.

What was interesting was that OFE2-based AI systems required fewer electronic parameters, making them lighter and less expensive to operate. Optical preprocessing was doing the hard work, and the AI models could focus on learning and interpretation. That synergy could potentially cut down on power consumption and hardware costs, while also speeding up performance.

That engine also tackled a financial trading application. Confronted with real-time market data, OFE2 processed price trends with speed and generated output signals that directly translated into buy or sell decisions. As the system operates at the speed of light, the decisions were made in virtual real time—a tremendous advantage in markets where every nanosecond counts.

What's notable about this optical engine isn't just speed—although it's that, as well—it's efficiency. Photonic systems consume less power than electronic ones, since photons don't get hot and can travel resistance-free. The Tsinghua researchers claim their device processes huge streams of data with very little energy loss, and preserves good signal integrity even under load.

Equally important is scalability. By adding more wavelength channels, the system can process more operations in parallel without a performance loss or additional latency. That versatility could make all-optical neural networks—a Holy Grail for many AI engineers—more of a reality than imagined.

While the optical engine is powerful, it won’t replace electronic processors overnight. Instead, future AI systems may use hybrid architectures, where photonic chips handle the lightning-fast feature extraction and electronics manage slower, precise tasks like training and control. This blend would merge the speed of light with the adaptability of silicon.

Scientists envision this research paving the way for fully trainable optical systems that can tune parameters directly with light, without utilizing electronics whatsoever. They're also developing optical memory and reconfigurable light filters, steps that would move computing entirely into the photonic domain.

For all the advances, there's still work to be done. It takes precise alignment and temperature control to manufacture these chips, and it is challenging to integrate several optical amplifiers onto one chip. Prices need to decrease as well before mass production can be achieved.

Nevertheless, the potential is too great to be disregarded. With ongoing innovation, optical computing may be supplementing—or even replacing—today's speediest silicon-based processors in the near future.

If it succeeds at scale, the technology could revolutionize the application of AI in nearly every domain. Self-driving cars could process visual data in real time. Robots performing surgery could react to unexpected complications as they arose. Trading systems could carry out trades more quickly and safely.

The energy efficiency alone would cut the carbon footprint of massive data centers that currently devour electricity.

By taking data processing to the speed of light, researchers are laying the foundation for AI that's not only faster but also greener and smarter.

Research findings are available online in the journal Advanced Photonics Nexus.

Like these kind of feel good stories? Get The Brighter Side of News' newsletter.