NeuralStackUniversity of Pennsylvania researchers have demonstrated a way to perform optical signal switching with exciton-polaritons, hybrid light-matter particles that could help move some AI computing tasks beyond conventional electronics. The work points toward chips that use light not only to carry information, but also to perform operations needed for computation.
Exciton-polaritons form when photons strongly couple with electrons in an atomically thin semiconductor. The resulting quasiparticles combine light’s speed and low-loss information transport with matter’s ability to interact, giving researchers a way to control optical signals more effectively.
That combination addresses a longstanding limitation of using light for computing. Photons can carry information quickly over long distances, but their lack of charge and mass makes them difficult to use for the signal-switching logic on which computers depend.
Electronic systems face a different set of constraints. Electrons generate heat and encounter resistance as they move through materials, challenges that become more significant as AI hardware handles increasingly large volumes of data.
The Penn team demonstrated all-light switching using about 4 quadrillionths of a joule. The result is particularly relevant to photonic AI chips, which can already use light for certain calculations but commonly need to convert signals back into electronic form for nonlinear activation steps and other decision-making operations.
Why this matters
Keeping those operations in the optical domain could reduce the energy and delay associated with repeated conversions between light and electricity. For AI systems, where data movement and computation place heavy demands on hardware, the ability to switch signals optically could make photonic architectures more practical.
The approach could also allow future chips to process information directly from cameras without repeatedly translating optical input into electrical signals. Researchers say it may eventually support basic quantum computing functions as well, although the technology still depends on successful scaling beyond the demonstrated system.
The research was led by Penn physicist Bo Zhen, with Li He, Zhi Wang, and Bumho Kim among the study's authors. Published in Physical Review Letters, the work represents an early step toward integrating light-based switching into faster, more energy-efficient AI hardware.