Swinburne University of Technology Researchers Demonstrate Optical Neuromorphic Processor Using Micro-Comb Photonic Chip
In January 2021, a Swinburne University of Technology-led international team published results in Nature demonstrating an optical neuromorphic processor built on a photonic micro-comb chip, achieving a processing speed of 10 TOPS from a single integrated device and performing image classification tasks at high speed.
Background
Electronic processors handle neural-network computations by moving data between memory and processing cores. At high throughput, this becomes a bottleneck: the wires carrying that data can only move so much information at once, and every operation consumes power. Researchers had known for decades that light could carry information far more efficiently than electrons, but turning that idea into a working neural-network accelerator was a different matter.
Photonic chips had been used for communications, where the goal is simply to transmit data quickly from one place to another. Using them to perform computation is harder. A neural network needs to multiply many numbers together simultaneously, weight them, and sum the results. Doing that optically, in a way that maps onto real machine-learning tasks, had not been achieved at any speed worth comparing to electronics.
One candidate approach used a micro-comb, a device that generates hundreds of separate laser lines at different wavelengths from a single chip. Each wavelength channel can carry a different value at the same time, which means a micro-comb can represent the weights of a neural network in parallel, spread across light rather than stored in electronic memory. The question was whether this could be made to work at a scale and speed that mattered.
What happened
In January 2021, a team led by Xingyuan Xu at Swinburne University of Technology, working with colleagues from RMIT University, Monash University, INRS Énergie Matériaux Télécommunications, and City University of Hong Kong, published their results in Nature. They had built a photonic convolutional accelerator, a chip that performs the sliding-window multiply-and-sum operations at the heart of a convolutional neural network, which is a type of network that processes data by scanning it in small overlapping patches rather than all at once.
The processor used an integrated photonic micro-comb, generating a frequency comb in the infrared, to encode neural-network weights across 49 parallel wavelength channels. Because all 49 channels operate simultaneously, the chip reached a processing speed of 11 TOPS, or tera-operations per second. The team used this to run image classification tasks on two standard datasets, CIFAR-10 and MNIST, reaching classification accuracies of 88.51% and 98.83% respectively.
That speed came from a single integrated device. The comparison that matters is not with the raw peak figures of large electronic processor arrays, but with what a single chip can do: no bank of parallel hardware, no external memory system moving weights back and forth. The result showed that encoding computation in wavelengths of light, rather than in voltages, could reach speeds that electronic integrated circuits of comparable physical scale could not match for this class of workload.
Why it mattered
The work showed that photonic integrated circuits could perform neuromorphic inference at speeds and energy profiles unattainable by conventional electronic processors, pointing toward a viable hardware path for optical neural-network acceleration. By encoding neural-network weights across the parallel wavelength channels of a micro-comb, the team demonstrated that a single chip could replace large banks of parallel electronic processors for certain inference workloads. The result strengthened the case for optical computing as a complementary substrate to silicon in AI hardware, particularly for latency-critical and bandwidth-hungry applications.
People
Xingyuan Xu, Mengxi Tan, Bill Corcoran, Jiayang Wu, Andreas Boes, Thach Nguyen, Sai T Chu, Brent E Little, Roberto Morandotti, Arnan Mitchell, David J Moss
Organisations
Swinburne University of Technology, RMIT University, Inrs Energie Materiaux Telecommunications, City University of Hong Kong, Monash University
Sources
- 11 TOPS photonic convolutional accelerator for optical neural networks.Nature.Primary source
- Swinburne-led research team demonstrates world's fastest optical neuromorphic processor.Swinburne University of Technology.Official
- 11 TOPS photonic convolutional accelerator for optical neural networks (Nature abstract page).Nature.Secondary
Cite this page
AI Achievements. (2021). Swinburne University of Technology Researchers Demonstrate Optical Neuromorphic Processor Using Micro-Comb Photonic Chip. Retrieved 2026-08-22, from https://achievements.ai/milestone/swinburne-university-neuromorphic-processor
@misc{achievements_swinburne_university_neuromorphic_processor,
title = {Swinburne University of Technology Researchers Demonstrate Optical Neuromorphic Processor Using Micro-Comb Photonic Chip},
author = {{AI Achievements}},
year = {2021},
url = {https://achievements.ai/milestone/swinburne-university-neuromorphic-processor}
}