Tech

Penn Researchers Build Light-Based Particles to Cut AI Power Use

June 18, 2026 · From the archive, written in the Professor's voice before our October 2026 redesign from sciencedaily.com

What happened

University of Pennsylvania scientists created hybrid light-matter quasiparticles called polaritons that perform matrix multiplications with photons instead of electrons. The method replaces certain electronic operations with optical computing, reducing energy per calculation by orders of magnitude. Early tests show inference speeds increase while heat output drops sharply.

Why it matters

This technique teaches practitioners to look beyond silicon for matrix-heavy workloads and consider optical or analog accelerators. Teams should audit which layers of their models could run on emerging photonic hardware. The shift reframes efficiency as a hardware choice rather than just model compression.

Who's doing it

The Penn electrical engineering group reported polariton-based accelerators running small neural networks at 10x lower power than GPU equivalents in lab benchmarks. Their results appear in recent ScienceDaily coverage and include measured latency figures.

Try it

1. Read the polariton paper linked from the ScienceDaily article at https://www.sciencedaily.com/news/computers_math/artificial_intelligence/.
2. Install the open-source Photontorch library and run their example photonic circuit simulation on a simple linear layer.
3. Compare energy estimates from the simulator against a standard PyTorch CPU run to quantify the difference.

Read the original at sciencedaily.com