A research team has developed a photonic device capable of performing matrix-vector multiplication in a non-volatile, compact and broadband format, aimed at accelerating neuromorphic computing workloads. The matrix-vector operation is at the core of artificial neural networks, and moving it into the optical domain promises to reduce energy consumption and increase speed compared with purely electronic implementations.

The work addresses a recurring limitation of photonic accelerators: the need to maintain the network weight configuration through continuous energy consumption, or else sacrifice bandwidth and chip footprint. The presented architecture combines non-volatile weight storage with a compact format that allows operation on optical signals across a wide spectral range, without requiring constant thermal or electrical feedback.

In terms of operation, the device encodes the input vector in the amplitude or phase of multiple optical modes and makes it interact with a weight matrix stored within the photonic circuit itself. The multiplication occurs intrinsically during light propagation, and the result is read directly in the optical domain, avoiding intermediate conversions that penalise performance in hybrid electronic-photonic systems.

The relevance of the advance lies in its potential for integration into existing integrated photonic platforms, bringing optical inference accelerators closer to practical applications in edge artificial intelligence and data centres. The authors point out that the combination of non-volatility, compactness and broadband operation lays the groundwork for scaling operations per second and per watt. As next steps, the community expects to see these blocks integrated into larger-scale systems and the characterisation of their arithmetic precision in real deep-learning tasks.