Forget electrons, this breakthrough uses light-matter particles to power AI

ENIAC, developed by Penn researchers J. Presper Eckert and John Mauchly, helped launch the modern computing era by using streams of electrons to solve complex mathematical problems. That same electronic approach still powers today's computers, smartphones, and AI systems. But as artificial intelligence grows more demanding, the limits of electron-based hardware are becoming harder to ignore.

Why Electrons Are Reaching Their Limits

Electrons carry an electrical charge, which creates several challenges inside modern computer chips. As they move through materials, they generate heat and face resistance that wastes energy. Those problems become even more difficult as chips grow more complex and process enormous amounts of data for AI applications.

Researchers led by Penn physicist Bo Zhen in the School of Arts & Sciences believe photons, the particles that make up light, could help solve some of these issues.

"Because they are charge-neutral and have zero rest mass, photons can carry information quickly over long distances with minimal loss, dominating communications technology," explains Li He, co-first author of a paper published in Physical Review Letters and a former postdoctoral researcher in the Zhen Lab. "But that neutrality means they barely interact with their environment, making them bad at the sort of signal-switching logic that computers depend on."

In other words, light is excellent for carrying information quickly and efficiently, but it struggles with the switching operations needed for computing.

Combining Light and Matter for AI Computing

To overcome that problem, Zhen's team developed a special quasiparticle called an exciton-polariton. The particle forms when photons are strongly linked with electrons inside an atomically thin semiconductor material. This combination allows light to interact much more effectively, making it capable of performing the signal switching required for computing tasks.

The breakthrough could be especially important for artificial intelligence systems, which consume enormous amounts of power.

Many experimental photonic AI chips already use light to handle certain calculations at high speed. However, when these systems need to perform nonlinear activation steps, such as decision-making operations, they usually must convert light signals back into electronic ones. That conversion slows the process and increases energy use, reducing the benefits of photonic computing.

Using exciton-polaritons, the Penn researchers demonstrated all-light switching while using only about 4 quadrillionths of a joule of energy. That amount is extraordinarily small, far below the energy needed to briefly power a tiny LED light.

Toward Faster and More Efficient AI Chips

If the technology can be successfully scaled, it could lead to photonic chips capable of processing information directly from cameras without repeated conversions between light and electricity. The approach could also lower the massive energy demands of large AI systems and potentially support basic quantum computing functions on future chips.

Bo Zhen is the Jin K. Lee Presidential Associate Professor in the Department of Physics and Astronomy in the School of Arts & Sciences at the University of Pennsylvania.

Li He was a postdoctoral researcher in the Zhen Lab in Penn Arts & Sciences. He is currently an assistant professor at Montana State University.

Additional authors on the study include Zhi Wang and Bumho Kim from the University of Pennsylvania's School of Arts & Sciences.

The research was supported by the US Office of Naval Research (N00014-20-1-2325 and N00014-21-1-2703) and the Sloan Foundation.