# Scalable photonic reservoir computing for parallel machine learning tasks

## TL;DR

Researchers developed a photonic computing device that can perform multiple machine learning tasks at high speed and low energy.

## Contribution

A tunable photonic reservoir computing device using a nonlinear amplifying loop mirror for parallel machine learning tasks.

## Key findings

- The device achieves a throughput of 20 tera-operations-per-second.
- It operates with an energy efficiency of 4.4 fJ per operation.

## Abstract

Neuromorphic photonics enables brain-inspired information processing with higher bandwidth and lower energy consumption than traditional electronics, addressing the growing computational demands of the Internet of Things, cloud services, and edge computing. However, even current state-of-the-art electronic and photonic platforms are incapable of delivering the scalable throughput, multitasking processing, and energy efficiency required by these applications. Here, we demonstrate a tunable photonic reservoir computing device based on a nonlinear amplifying loop mirror (NALM), leveraging a time-delayed, single-unit, all-optical architecture. By combining dense temporal encoding with wavelength-division multiplexing, the system supports concurrent multitasking across independent data channels, enabling scalable computational performance without additional hardware complexity. Experiments and theoretical validation on classification and prediction benchmarks demonstrate the device’s performance, achieving a throughput of 20 tera-operations-per-second and an energy efficiency of 4.4 fJ per operation. These results highlight a promising path towards reconfigurable, compact, and high-performance photonic processors for real-time intelligent applications.

Neuromorphic computing processes data faster and with less energy than electronics. Here, authors demonstrate a reconfigurable photonic reservoir computer that performs multiple machine learning tasks in parallel at ultrafast rates while using extremely low energy per operation.