Penn Team Builds Hybrid Light-Matter Particle to Accelerate AI Inference
Engineers at the University of Pennsylvania coupled a silicon photonic resonator with an electronic neuron, creating a polariton that performs matrix-vector multiplication at 50 GHz using 0.3 picojoules per operation. The device handled 1024-by-1024 multiplications in a single clock cycle with 4-bit precision.
You begin to treat part of your model as an optical co-processor rather than another GPU thread. Latency budgets shift from milliseconds on CUDA cores to nanoseconds on-chip, changing which layers you decide to keep electronic and which you offload to photonics.
Professor Liang Feng’s lab at Penn demonstrated the polariton chip on a 4-by-4 MNIST classifier and reported 98.7 percent accuracy at 0.8 W total system power, published in Nature Photonics, May 2026.
Step 1: request access to the Penn Polariton SDK at photonics.seas.upenn.edu/polariton-sdk. Step 2: wrap your PyTorch linear layer with the PolaritonLinear class and set precision to 4 bits. Step 3: run the included latency_test.py to measure a drop from 1.2 ms to 18 ns per inference on their evaluation board.