Penn team advances light-based AI chips

University of Pennsylvania researchers have developed exciton-polaritons, hybrid light-matter quasiparticles formed by strongly linking photons with electrons inside an atomically thin semiconductor. The approach targets a core weakness of photonic computing: light can carry information quickly and efficiently, but its neutral nature makes it poor at the signal-switching logic computers require.

The team demonstrated all-light switching using only about 4 quadrillionths of a joule of energy. By enabling nonlinear activation steps without converting light signals back into electronic ones, the work could reduce delays and energy losses in experimental photonic AI chips.

If the method can be scaled, photonic chips could process information directly from cameras and avoid repeated conversions between light and electricity. The researchers say the approach could help lower the energy demands of large AI systems and may support basic quantum computing functions on future chips.