Penn Researchers Create 4-Quadrillionth-Joule Light-Matter Switch for AI Chips
Updated
Updated · ScienceDaily · May 19
Penn Researchers Create 4-Quadrillionth-Joule Light-Matter Switch for AI Chips
4 articles · Updated · ScienceDaily · May 19
University of Pennsylvania scientists built an exciton-polariton device that performs all-light switching at about 4 quadrillionths of a joule, a key step toward lower-power AI computing.
The hybrid quasiparticle couples photons with electrons in an atomically thin semiconductor, aiming to solve a core photonics problem: light moves data efficiently but normally cannot handle the nonlinear switching computers need.
That matters for AI because photonic chips already speed some calculations, yet still convert signals back into electronics for activation and decision steps, adding delay and energy loss.
If scaled, the approach could enable chips that process camera data directly in light, cut the heavy power demands of large AI systems, and support some basic quantum-computing functions.
Can new light-powered chips scale fast enough to avert the AI energy crisis?
How will this ultra-efficient AI reshape the future of autonomous military technology?
Breakthrough All-Optical Switch from UPenn Sets New Benchmark for Energy-Efficient AI Hardware
Overview
In 2024, the University of Pennsylvania introduced a breakthrough in AI hardware by developing a novel Optical Neural Network (ONN) system that acts as an all-optical switch for AI chips. This innovation addresses a major bottleneck in photonic AI hardware by allowing optical signals to be controlled directly on a semiconductor chip, eliminating the need to convert light signals into electronic ones for processing. The ONN system works by shining a laser onto specific parts of the chip, dynamically changing its optical properties and enabling flexible, reprogrammable pathways for AI computations. This advancement promises faster, more efficient, and adaptable AI hardware.