Matrix Multiplication Unit Scalability Investigation for InP SOA-based Photonic Deep Neural Network

B. Shi; N. Calabretta; R. Stabile

doi:10.1109/SUM48678.2020.9161056

Matrix Multiplication Unit Scalability Investigation for InP SOA-based Photonic Deep Neural Network

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Academic › peer-review

Abstract

We investigate the scalability of a SOA-based photonic DNN with WDM inputs. For a 3-layer DNN, a quasi-linear dependence between the prediction accuracy and errors/layer is found for a NRMSE <0.09. Optimized passive losses can enable synaptic function at half of the power consumption.

Original language	English
Title of host publication	2020 IEEE Photonics Society Summer Topical Meeting Series, SUM 2020 - Proceedings
Publisher	Institute of Electrical and Electronics Engineers
Number of pages	2
ISBN (Electronic)	978-1-7281-5887-7
DOIs	https://doi.org/10.1109/SUM48678.2020.9161056
Publication status	Published - Jul 2020
Event	2020 IEEE Photonics Society Summer Topical Meeting Series, SUM 2020 - Cabo San Lucas, Mexico Duration: 13 Jul 2020 → 15 Jul 2020

Conference

Conference	2020 IEEE Photonics Society Summer Topical Meeting Series, SUM 2020
Country/Territory	Mexico
City	Cabo San Lucas
Period	13/07/20 → 15/07/20

Funding

This work is financially supported by the Netherlands Organization of Scientific Research (NWO) under the Gravitation program, (Zwaartekracht

Keywords

Artificial neural networks
Image classification
Photonic integrated circuits
Semiconductor optical amplifiers
Simulation

Access to Document

10.1109/SUM48678.2020.9161056

Cite this

@inproceedings{b8421244daac47359db13a3f33721ada,

title = "Matrix Multiplication Unit Scalability Investigation for InP SOA-based Photonic Deep Neural Network",

abstract = "We investigate the scalability of a SOA-based photonic DNN with WDM inputs. For a 3-layer DNN, a quasi-linear dependence between the prediction accuracy and errors/layer is found for a NRMSE <0.09. Optimized passive losses can enable synaptic function at half of the power consumption.",

keywords = "Artificial neural networks, Image classification, Photonic integrated circuits, Semiconductor optical amplifiers, Simulation",

author = "B. Shi and N. Calabretta and R. Stabile",

year = "2020",

month = jul,

doi = "10.1109/SUM48678.2020.9161056",

language = "English",

booktitle = "2020 IEEE Photonics Society Summer Topical Meeting Series, SUM 2020 - Proceedings",

publisher = "Institute of Electrical and Electronics Engineers",

address = "United States",

note = "2020 IEEE Photonics Society Summer Topical Meeting Series, SUM 2020 ; Conference date: 13-07-2020 Through 15-07-2020",

}

Shi, B , Calabretta, N & Stabile, R 2020, Matrix Multiplication Unit Scalability Investigation for InP SOA-based Photonic Deep Neural Network. in 2020 IEEE Photonics Society Summer Topical Meeting Series, SUM 2020 - Proceedings., 9161056, Institute of Electrical and Electronics Engineers, 2020 IEEE Photonics Society Summer Topical Meeting Series, SUM 2020, Cabo San Lucas, Mexico, 13/07/20. https://doi.org/10.1109/SUM48678.2020.9161056

Matrix Multiplication Unit Scalability Investigation for InP SOA-based Photonic Deep Neural Network. / Shi, B.; Calabretta, N.; Stabile, R.
2020 IEEE Photonics Society Summer Topical Meeting Series, SUM 2020 - Proceedings. Institute of Electrical and Electronics Engineers, 2020. 9161056.

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Academic › peer-review

TY - GEN

T1 - Matrix Multiplication Unit Scalability Investigation for InP SOA-based Photonic Deep Neural Network

AU - Shi, B.

AU - Calabretta, N.

AU - Stabile, R.

PY - 2020/7

Y1 - 2020/7

N2 - We investigate the scalability of a SOA-based photonic DNN with WDM inputs. For a 3-layer DNN, a quasi-linear dependence between the prediction accuracy and errors/layer is found for a NRMSE <0.09. Optimized passive losses can enable synaptic function at half of the power consumption.

AB - We investigate the scalability of a SOA-based photonic DNN with WDM inputs. For a 3-layer DNN, a quasi-linear dependence between the prediction accuracy and errors/layer is found for a NRMSE <0.09. Optimized passive losses can enable synaptic function at half of the power consumption.

KW - Artificial neural networks

KW - Image classification

KW - Photonic integrated circuits

KW - Semiconductor optical amplifiers

KW - Simulation

UR - https://www.scopus.com/pages/publications/85090143471

U2 - 10.1109/SUM48678.2020.9161056

DO - 10.1109/SUM48678.2020.9161056

M3 - Conference contribution

AN - SCOPUS:85090143471

BT - 2020 IEEE Photonics Society Summer Topical Meeting Series, SUM 2020 - Proceedings

PB - Institute of Electrical and Electronics Engineers

T2 - 2020 IEEE Photonics Society Summer Topical Meeting Series, SUM 2020

Y2 - 13 July 2020 through 15 July 2020

ER -

Matrix Multiplication Unit Scalability Investigation for InP SOA-based Photonic Deep Neural Network

Abstract

Conference

Funding

Keywords

Access to Document

Other files and links

Fingerprint

Cite this