Tech

Hybrid light-matter quasiparticles promise faster, cooler AI accelerators

June 10, 2026 · From the archive, written in the Professor's voice before our October 2026 redesign from sciencedaily.com

What happened

Penn physicists coupled photons to excitons inside a 2-D perovskite microcavity to form polaritons whose collective spin processes matrix multiplications at 1.2 picoseconds per operation. Early prototypes show 40 times lower heat dissipation than equivalent electronic tensor cores.

Why it matters

The demonstration shows that moving part of the computation from electrons to photons removes a fundamental speed limit. You begin to view optical interconnects and photonic co-processors as practical options rather than laboratory curiosities.

Who's doing it

The photonic-AI startup Lumiphase integrated the Penn polariton gates into a prototype inference card that delivered 9,800 images per second on ResNet-50 while drawing 18 watts, versus 220 watts for an NVIDIA A100 running the same workload.

Try it

1. Download the open polariton simulator from https://github.com/upenn-photonics/polariton-torch.
2. Convert a single linear layer in your model to PolaritonLinear(in_features, out_features).
3. Execute a 1,000-sample inference run; measure a 35-fold reduction in energy per image and latency below 2 milliseconds.

Read the original at sciencedaily.com