Penn researchers demonstrated all-optical signal switching at roughly 4 femtojoules per operation, published May 18 in Physical Review Letters, volume 136, issue 14.

The demonstration originated in the laboratory of Bo Zhen, Jin K. Lee Presidential Associate Professor in the Department of Physics and Astronomy at the University of Pennsylvania's School of Arts & Sciences. The paper is titled "Strongly Nonlinear Nanocavity Exciton Polaritons in Gate-Tunable Monolayer Semiconductors."

Co-first author Li He set out the starting premise. "Because they are charge-neutral and have zero rest mass, photons can carry information quickly over long distances with minimal loss, dominating communications technology," He said. He worked as a postdoctoral researcher in the Zhen Lab during the study.

He now serves as an assistant professor at Montana State University.

What did the team actually build?

The group engineered a quasiparticle called an exciton-polariton. The particle forms when photons couple strongly to electrons inside an atomically thin, gate-tunable monolayer semiconductor. The hybrid state gives light enough interaction strength to perform nonlinear switching.

That gap matters for AI accelerators. Experimental photonic AI chips already perform linear operations at high speed. They stall at the nonlinear activation step, the decision-making logic in every neural network layer. Prototypes convert the optical signal back to an electrical one for that step. That conversion erodes the energy and latency advantage of light-based computing.

Electrons carry an electrical charge. As they move through materials, they generate heat and face resistance that wastes energy. Those losses scale poorly with the data volumes AI workloads demand. Photons, by contrast, lose little energy over distance and produce minimal heat.

How small is the energy budget?

The prototype consumed roughly 4 quadrillionths of a joule for the switching operation. The figure approximates the energy needed to flash a tiny LED for a fraction of a second. The photon budget is the headline number of the demonstration.

He framed the trade-off that limits bare photon logic. "That neutrality means they barely interact with their environment, making them bad at the sort of signal-switching logic that computers depend on," He added. The exciton-polariton approach restores that interaction.

What could change for AI hardware?

The platform could enable photonic processors that ingest image data directly from cameras. Repeated optical-to-electrical conversion at sensor interfaces would no longer be required. The authors project lower energy consumption across large AI inference systems. They also flag possible support for primitive quantum-computing operations on future photonic chips.

Commercial translation is not imminent. The achievement is a device-level demonstration in a single monolayer. Scaling to wafer-fab-compatible process nodes and high-yield manufacturing remains the open question. Foundry or back-end-of-line integration has not been demonstrated.

Who paid for the work?

The US Office of Naval Research supported the research under grants N00014-20-1-2325 and N00014-21-1-2703. The Sloan Foundation provided additional funding. Co-authors Zhi Wang and Bumho Kim, both of Penn Arts & Sciences, completed the authorship list. The full citation is DOI 10.1103/gc15-qsvf.

Wider context connects to the same campus that built ENIAC in 1945. The University of Pennsylvania housed the researchers who built the world's first general-purpose electronic computer. Eighty years on, Penn researchers are again rethinking how information moves through a processor. The bond between computing and electrons is loosening, deliberately.

Whether exciton-polariton switching survives process-scaling and yield tests will determine its place in the next wave of low-power AI accelerators, and whether photonic computing finally crosses from optical interconnect into optical logic.