NeuralStackResearchers at the University of Pennsylvania have developed a hybrid light-matter particle that could help move some AI computing tasks away from conventional electronics. Called an exciton-polariton, the quasiparticle forms when photons strongly couple with electrons in an atomically thin semiconductor material.
The result combines light’s ability to carry information quickly with matter’s ability to interact and switch signals. That combination allowed the Penn team, led by physicist Bo Zhen, to demonstrate signal switching using light alone.
Modern computers still rely on electrons, the foundation of electronic computing since the 1940s. As AI systems process larger volumes of data, however, electrical resistance and heat increasingly constrain performance and energy efficiency.
Photons offer an attractive alternative because they are charge-neutral and have no rest mass, allowing them to carry information rapidly with minimal loss. Their weakness is that they interact only weakly with their surroundings, making them poorly suited to the signal-switching logic required for many computing operations.
Photonic AI systems already use light for some high-speed calculations, but nonlinear activation steps—operations that help systems make decisions—often require converting optical signals back into electronic ones. Those conversions add delay and energy costs, limiting the advantages of light-based processing.
Why this matters
The Penn demonstration suggests that exciton-polaritons could let photonic systems perform more of their processing without leaving the optical domain. The researchers reported all-light switching at about 4 quadrillionths of a joule, an exceptionally small amount of energy.
If the method can be scaled, future photonic chips could process information directly from cameras without repeatedly converting light into electricity. That could reduce the energy demands of large AI systems while improving the speed of data movement and computation.
The platform may also have uses beyond AI, including basic quantum computing functions on future chips. The work, published in Physical Review Letters, was conducted by Bo Zhen, Li He, Zhi Wang, and Bumho Kim at the University of Pennsylvania, with Li He now an assistant professor at Montana State University.