Hybrid light-matter quasiparticles slash AI energy use at Penn
Researchers at the University of Pennsylvania created polaritons, hybrid light-matter quasiparticles, to replace selected electronic logic gates. The device performed matrix multiplications at 10 times lower energy per operation than standard CMOS circuits. The team measured a 40 percent reduction in heat while maintaining 92 percent accuracy on a 784-by-128 neural network layer.
You stop treating compute as an unlimited resource and start designing workflows that offload linear algebra to optical paths. This reframes model training budgets around energy per FLOP instead of raw throughput. The practical result is earlier decisions about when to keep tasks on GPUs and when to move them to emerging photonic accelerators.
The University of Pennsylvania Excitonics Lab ran the polariton chip on a ResNet-18 backbone and cut inference energy from 3.2 joules to 1.9 joules per image on CIFAR-10. The group published device parameters and open-sourced the simulation scripts at github.com/UPenn-Excitonics/polariton-sim.
Step 1: Visit https://github.com/UPenn-Excitonics/polariton-sim and clone the repository. Step 2: Run the included Jupyter notebook polariton_benchmark.ipynb with your own 784-by-128 weight matrix to obtain estimated energy per operation. Step 3: Compare the notebook output against your current GPU benchmark; if the ratio favors polaritons by more than 3x, schedule a hardware pilot with the lab contact listed in the README.