⚡ AI at the speed of light: it's possible

Nearly eighty years ago, the ENIAC marked the beginning of the electronic era. Today, our computers still rely on the same principle: moving electrons. Yet this technology is reaching its limits, especially for artificial intelligence.

Researchers at the University of Pennsylvania are proposing an alternative: quasiparticles mixing light and matter, capable of performing calculations without electrons. These particles, called exciton-polaritons, could replace electrons in certain tasks, offering speed and low energy consumption.


Electrons carry an electric charge, which generates heat and resistance when they move in chips. The larger AI systems become, the more these problems worsen. Heat dissipation becomes a major issue, and wasted energy hampers performance. Researchers are therefore looking for alternatives to avoid these thermal and energy bottlenecks.

Photons, particles of light, seem ideal: they have no charge and move quickly without loss. But they interact very little with their environment, which makes them poor for logic switching operations, essential in computing. As Li He, a former researcher at the Zhen Lab, explains, photons excel at transporting information over long distances, but their neutrality prevents them from making the kind of binary decisions that computers require.

Exciton-polaritons are quasiparticles formed by coupling photons with electrons in an ultrathin semiconductor. This union gives rise to a hybrid object that combines the speed of light and the interaction capability of matter. Researchers have succeeded in creating these particles in a nanocavity, paving the way for all-optical logic operations without recourse to electrons.

For AI, this advance is significant. Current photonic chips use light, but must convert signals into electricity for nonlinear steps, such as decisions. These repeated conversions slow things down and consume energy. With exciton-polaritons, it is possible to perform these operations directly with light, eliminating costly back-and-forth between photons and electrons.

The team demonstrated all-optical switching using only 4 quadrillionths of a joule, a tiny amount of energy. This paves the way for much more efficient chips. If the technology scales, it could allow processing light directly without endless conversions, drastically reducing the energy consumption of AI systems.

In the long run, these photonic chips could significantly reduce the energy demand of large AI models. They could even support basic quantum computing functions. The researchers, led by Bo Zhen, continue to explore the possibilities, but the potential is immense. Their results were published in Physical Review Letters.

What is an exciton-polariton?

An exciton-polariton is a hybrid quasiparticle that arises from the coupling between a photon and an exciton (an electron-hole pair) in a semiconductor. This mixture gives the particle unique properties: it behaves both as a light wave, therefore very fast, and as a material particle capable of interactions.

These quasiparticles generally form in nanoscale optical cavities, where light is confined. Choosing an atomically thin semiconductor, such as molybdenum disulfide, allows fine control of properties via an electrical voltage.

The interest for computing is immense: exciton-polaritons allow performing logic operations at the speed of light. They could thus replace electrons in certain key functions of processors.