Hybrid light-matter quasiparticles promise faster, cooler AI chips

Penn researchers coupled photons with excitons in a 2-D perovskite microcavity to form polaritons that perform matrix multiplies at the speed of light. Their prototype executes a 1024-by-1024 multiply in 120 femtoseconds while drawing 40 femtojoules per operation, two orders of magnitude below electronic SRAM. The device is fabricated with standard lithography on a silicon substrate.

You begin evaluating workloads for optical versus electronic paths instead of assuming all computation stays on transistors. This reframes model architecture choices around latency and heat budgets rather than FLOPs alone.

Professor Ritesh Agarwal's lab at the University of Pennsylvania ran a 4-by-4 polariton array for 72 hours continuous inference at 1.2 mW total system power, publishing the trace data and mask layouts under an open-source hardware license.