Penn researchers fuse photons and excitons to slash AI power draw

University of Pennsylvania physicists created a hybrid light-matter quasiparticle called a polariton inside a microcavity containing a two-dimensional perovskite. When driven by a weak laser, the polariton lattice performed matrix multiplications at room temperature while consuming microwatts instead of the watts required by digital GPUs. The device ran a simple neural-network inference loop with 150 times lower energy per operation than an A100 GPU running the same workload.

Engineers see that analog optical accelerators can replace selected digital layers, so the new habit is to profile which sublayers tolerate low-precision analog math and offload them first.

The Penn Nanophotonics lab has taped out a 64-by-64 polariton array chip; preliminary measurements show 0.3 femtojoules per multiply-accumulate versus 30 femtojoules on an NVIDIA A100.