Penn researchers fuse photons and excitons to slash AI energy costs
A team at the University of Pennsylvania built hybrid light-matter quasiparticles called polaritons inside a specially engineered microcavity. These polaritons replace some electronic switching steps with optical interference, cutting both latency and power draw during matrix multiplications common in transformer training. Early tests showed inference speeds rising by a factor of ten while energy per operation fell below one femtojoule.
You stop treating every AI workload as an electron-only problem and start asking which layers can move to light-based gates. This reframes hardware selection from GPU versus TPU to GPU versus photonic co-processor. The workflow change is to profile matrix size and sparsity first, then offload eligible subgraphs to optical accelerators instead of defaulting to more silicon.
The University of Pennsylvania Excitonics Lab has taped out a 64-by-64 polariton array and reported 9.8 times faster inference on a 1.3-billion-parameter language model subset with 14 times lower energy than an equivalent A100 run.
Step 1: Visit the Penn Excitonics Lab publications page at https://www.seas.upenn.edu/~excitonics/ and download their open-source polariton simulation notebook. Step 2: Run the included Jupyter notebook on a laptop with an NVIDIA GPU to model a 16-by-16 polariton mesh and record latency and power estimates. Step 3: Compare those numbers against your current PyTorch inference script on the same matrix size; if optical estimates beat your baseline by more than 3 times, schedule a meeting with your hardware team to evaluate photonic co-processors.