Penn physicists fuse photons and electrons into hybrid quasiparticles to slash AI energy costs

Researchers at the University of Pennsylvania created polaritons, hybrid light-matter particles, that replace selected electronic gates in neural network accelerators. The device performed matrix multiplications at 40 femtojoules per operation versus 1 picojoule for standard CMOS, cutting energy per inference by roughly 25 times. The team used a gallium arsenide microcavity coupled to a monolayer of transition-metal dichalcogenide and measured coherent polariton propagation at room temperature.

The work shows that replacing arithmetic units with analog optical degrees of freedom can bypass the von Neumann bottleneck. Practitioners should audit which layers of their models perform dense linear algebra and test whether those layers can migrate to photonic substrates before scaling cluster size.

Professor Ritesh Agarwal's group at Penn has fabricated and tested the polariton gates; their 2025 arXiv preprint reports inference latency of 12 nanoseconds on a 784 by 256 fully connected layer with measured energy of 0.9 picojoules per forward pass.