Engineers at the University of Pennsylvania have created a new kind of computer chip that uses light instead of electricity to train artificial intelligence (AI) systems.
This breakthrough could dramatically speed up AI training while using much less energy.
Unlike most AI chips today that rely on electronic signals, this new chip is photonic, meaning it uses beams of light to do its work.
What makes this development so important is that the chip can now handle a type of math called “nonlinear functions,” which are essential for teaching AI systems to learn and make decisions.
This is the first time this has been done using only light.
Professor Liang Feng, who led the study, explains that nonlinear functions are what allow deep neural networks—complex layers of connected units that mimic how our brains work—to actually learn.
Without nonlinearity, these networks would be limited to simple calculations and wouldn’t be able to perform intelligent tasks.
In the past, researchers could only use light to perform basic, linear math on chips. But this new chip changes that.
It’s made of a special material that reacts to light in a way that can be controlled. One beam of light, called the “signal” beam, carries the data.
Another beam, the “pump” light, shines from above and adjusts how the material behaves.
By changing this pump light, researchers can program the chip to act differently, allowing it to perform a variety of nonlinear functions—all with light.
This setup means the chip can be reprogrammed just by adjusting the light shining on it. It doesn’t need to be rebuilt or physically changed, which makes it incredibly flexible.
The researchers tested the chip on classic AI tasks and found that it performed just as well, or even better, than traditional electronic AI systems—while using far less power.
For example, they showed that just four optical connections on the chip could match the performance of 20 traditional electronic connections. And unlike older photonic chips that were fixed once built, this one starts blank and can be updated on the fly.
The team hopes that in the future, this light-based technology could train even more advanced AI systems, like those used in large language models, and dramatically cut energy use in data centers.
Feng says this could mark the beginning of a new era in computing—one where light, not electricity, powers the next generation of intelligent machines. It’s a fitting development at Penn, where the first digital computer, ENIAC, was created. This photonic chip might just be the modern version of that historic leap forward.
Source: University of Pennsylvania.
