Related Experiment Video
Updated: Jun 29, 2026
08:41
Lensfree On-chip Tomographic Microscopy Employing Multi-angle Illumination and Pixel Super-resolution
Published on: August 16, 2012
11.5K
Analog nanophotonic computing going practical: silicon photonic deep learning engines for tiled optical matrix
George Giamougiannis1, Apostolos Tsakyridis1, Miltiadis Moralis-Pegios1
1Department of Informatics, Center for Interdisciplinary Research & Innovation, Aristotle University of Thessaloniki, Thessaloniki, Greece.
Nanophotonics (Berlin, Germany)
|December 5, 2024
Summary
This study introduces a novel dynamic precision neural network (NN) inference method using silicon photonic processors. This approach accelerates deep neural network (DNN) computations by 55% without sacrificing accuracy.
Area of Science:
- Photonics
- Artificial Intelligence
- Computer Engineering
Background:
- Analog photonic computing offers high bandwidth and low power for deep neural networks (DNNs).
- Limitations in hardware size and component bit precision hinder photonic processors from outperforming digital ones.
- Existing photonic solutions face challenges in achieving high-speed, high-accuracy DNN inference.
Purpose of the Study:
- To propose and demonstrate a speed-optimized dynamic precision neural network (NN) inference system.
- To develop a theoretical model linking photonic neuron noise to bit precision requirements.
- To accelerate DNN inference on a silicon photonic processor using tiled matrix multiplication (TMM).
Main Methods:
- Experimental demonstration of dynamic precision NN inference on a silicon photonic processor.
- Implementation of tiled matrix multiplication (TMM) for optical computations up to 50 GHz.
- Application of a theoretical model to determine layer-specific bit precision requirements.
- Mixed-precision inference scheme adjusting operational compute rates per neural layer.
Main Results:
- A 55% reduction in execution time for linear operations was achieved compared to fixed-precision methods.
- Dynamic-precision NN inference was successfully demonstrated on a silicon coherent photonic neuron.
- The system maintained classification accuracy for the IRIS dataset.
- A theoretical model correlating noise figure with bit precision was established.
Conclusions:
- Dynamic precision NN inference on silicon photonic processors significantly speeds up DNN computations.
- The proposed method overcomes limitations of fixed-precision photonic hardware.
- This approach enables high-accuracy, high-speed AI acceleration using analog photonic computing.
