Penn researchers build hybrid light-matter particles to cut AI energy use
Researchers at the University of Pennsylvania created polaritons, hybrid particles that combine photons and excitons. These particles were used in an optical neural network to perform matrix multiplications at light speed with lower power draw than electronic chips. The work was published May 18, 2026.
This shows readers that optical computing can replace some GPU operations. It pushes users to track hardware beyond silicon and consider energy cost in model selection. The result is a workflow that weighs latency against watts when choosing inference platforms.
The University of Pennsylvania photonics lab demonstrated the polariton device on a 784 by 128 matrix multiply task. Their prototype used 40 percent less energy than an equivalent electronic accelerator while maintaining accuracy above 96 percent.
Step 1: Visit the Penn Electrical and Systems Engineering site at https://www.ese.upenn.edu and download the polariton simulation notebook. Step 2: Run the provided Jupyter notebook on a standard CPU to model a 784 by 128 matrix multiply and record energy estimates. Step 3: Compare the notebook output against a local PyTorch benchmark on the same matrix to quantify the projected energy savings.