A Flexible FPGA-based Inference Architecture for Pruned Deep Neural Networks

Thorbjörn Posewsky; Daniel Ziener

doi:10.1007/978-3-319-77610-1_23

A Flexible FPGA-based Inference Architecture for Pruned Deep Neural Networks

Thorbjörn Posewsky, Daniel Ziener

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

8 Citations (Scopus)

5 Downloads (Pure)

Abstract

In this paper, we present an architecture for embedded FPGA-based deep neural network inference which is able to handle pruned weight matrices. Pruning of weights and even entire neurons reduces the amount of data and calculations significantly, thus improving enormously the efficiency and performance of the neural network inference in embedded devices. By using an HLS approach, the architecture is easily extendable and highly configurable with a free choice of parameters like the number of MAC units or the used activation function. For large neural networks, our approach competes with at least comparable performance as state-of-the-art x86-based software implementations while only using 10% of the energy.

Original language	English
Title of host publication	Architecture of Computing Systems – ARCS 2018
Subtitle of host publication	31st International Conference, Braunschweig, Germany, April 9–12, 2018, Proceedings
Editors	Mladen Berekovic, Rainer Buchty, Heiko Hamann, Dirk Koch, Thilo Pionteck
Place of Publication	Braunschweig, Germany
Publisher	Springer
Pages	311-323
Number of pages	13
ISBN (Electronic)	978-3-319-77610-1
ISBN (Print)	978-3-319-77609-5
DOIs	https://doi.org/10.1007/978-3-319-77610-1_23
Publication status	Published - 1 May 2018
Externally published	Yes
Event	31st International Conference on Architecture of Computing Systems 2018 - Technical University of Braunschweig, Braunschweig, Germany Duration: 9 Apr 2018 → 12 Apr 2018 Conference number: 31 http://arcs2018.itec.kit.edu/

Publication series

Name	Lecture notes in computer science
Volume	10793

Conference

Conference	31st International Conference on Architecture of Computing Systems 2018
Abbreviated title	ARCS 2018
Country/Territory	Germany
City	Braunschweig
Period	9/04/18 → 12/04/18
Internet address	http://arcs2018.itec.kit.edu/

Access to Document

10.1007/978-3-319-77610-1_23

Cite this

Posewsky, T., & Ziener, D. (2018). A Flexible FPGA-based Inference Architecture for Pruned Deep Neural Networks. In M. Berekovic, R. Buchty, H. Hamann, D. Koch, & T. Pionteck (Eds.), Architecture of Computing Systems – ARCS 2018: 31st International Conference, Braunschweig, Germany, April 9–12, 2018, Proceedings (pp. 311-323). (Lecture notes in computer science; Vol. 10793). Springer. https://doi.org/10.1007/978-3-319-77610-1_23

Posewsky, Thorbjörn ; Ziener, Daniel. / A Flexible FPGA-based Inference Architecture for Pruned Deep Neural Networks. Architecture of Computing Systems – ARCS 2018: 31st International Conference, Braunschweig, Germany, April 9–12, 2018, Proceedings. editor / Mladen Berekovic ; Rainer Buchty ; Heiko Hamann ; Dirk Koch ; Thilo Pionteck. Braunschweig, Germany : Springer, 2018. pp. 311-323 (Lecture notes in computer science).

@inproceedings{035d30a1fa4a4d0194d48ed544a93c65,

title = "A Flexible FPGA-based Inference Architecture for Pruned Deep Neural Networks",

abstract = "In this paper, we present an architecture for embedded FPGA-based deep neural network inference which is able to handle pruned weight matrices. Pruning of weights and even entire neurons reduces the amount of data and calculations significantly, thus improving enormously the efficiency and performance of the neural network inference in embedded devices. By using an HLS approach, the architecture is easily extendable and highly configurable with a free choice of parameters like the number of MAC units or the used activation function. For large neural networks, our approach competes with at least comparable performance as state-of-the-art x86-based software implementations while only using 10\% of the energy.",

author = "Thorbj{\"o}rn Posewsky and Daniel Ziener",

year = "2018",

month = may,

day = "1",

doi = "10.1007/978-3-319-77610-1\_23",

language = "English",

isbn = "978-3-319-77609-5",

series = "Lecture notes in computer science",

publisher = "Springer",

pages = "311--323",

editor = "Mladen Berekovic and Rainer Buchty and Heiko Hamann and Dirk Koch and Thilo Pionteck",

booktitle = "Architecture of Computing Systems – ARCS 2018",

address = "Germany",

note = "31st International Conference on Architecture of Computing Systems 2018, ARCS 2018 ; Conference date: 09-04-2018 Through 12-04-2018",

url = "http://arcs2018.itec.kit.edu/",

}

Posewsky, T & Ziener, D 2018, A Flexible FPGA-based Inference Architecture for Pruned Deep Neural Networks. in M Berekovic, R Buchty, H Hamann, D Koch & T Pionteck (eds), Architecture of Computing Systems – ARCS 2018: 31st International Conference, Braunschweig, Germany, April 9–12, 2018, Proceedings. Lecture notes in computer science, vol. 10793, Springer, Braunschweig, Germany, pp. 311-323, 31st International Conference on Architecture of Computing Systems 2018, Braunschweig, Germany, 9/04/18. https://doi.org/10.1007/978-3-319-77610-1_23

A Flexible FPGA-based Inference Architecture for Pruned Deep Neural Networks. / Posewsky, Thorbjörn; Ziener, Daniel.
Architecture of Computing Systems – ARCS 2018: 31st International Conference, Braunschweig, Germany, April 9–12, 2018, Proceedings. ed. / Mladen Berekovic; Rainer Buchty; Heiko Hamann; Dirk Koch; Thilo Pionteck. Braunschweig, Germany: Springer, 2018. p. 311-323 (Lecture notes in computer science; Vol. 10793).

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

TY - GEN

T1 - A Flexible FPGA-based Inference Architecture for Pruned Deep Neural Networks

AU - Posewsky, Thorbjörn

AU - Ziener, Daniel

N1 - Conference code: 31

PY - 2018/5/1

Y1 - 2018/5/1

N2 - In this paper, we present an architecture for embedded FPGA-based deep neural network inference which is able to handle pruned weight matrices. Pruning of weights and even entire neurons reduces the amount of data and calculations significantly, thus improving enormously the efficiency and performance of the neural network inference in embedded devices. By using an HLS approach, the architecture is easily extendable and highly configurable with a free choice of parameters like the number of MAC units or the used activation function. For large neural networks, our approach competes with at least comparable performance as state-of-the-art x86-based software implementations while only using 10% of the energy.

AB - In this paper, we present an architecture for embedded FPGA-based deep neural network inference which is able to handle pruned weight matrices. Pruning of weights and even entire neurons reduces the amount of data and calculations significantly, thus improving enormously the efficiency and performance of the neural network inference in embedded devices. By using an HLS approach, the architecture is easily extendable and highly configurable with a free choice of parameters like the number of MAC units or the used activation function. For large neural networks, our approach competes with at least comparable performance as state-of-the-art x86-based software implementations while only using 10% of the energy.

U2 - 10.1007/978-3-319-77610-1_23

DO - 10.1007/978-3-319-77610-1_23

M3 - Conference contribution

SN - 978-3-319-77609-5

T3 - Lecture notes in computer science

SP - 311

EP - 323

BT - Architecture of Computing Systems – ARCS 2018

A2 - Berekovic, Mladen

A2 - Buchty, Rainer

A2 - Hamann, Heiko

A2 - Koch, Dirk

A2 - Pionteck, Thilo

PB - Springer

CY - Braunschweig, Germany

T2 - 31st International Conference on Architecture of Computing Systems 2018

Y2 - 9 April 2018 through 12 April 2018

ER -

Posewsky T, Ziener D. A Flexible FPGA-based Inference Architecture for Pruned Deep Neural Networks. In Berekovic M, Buchty R, Hamann H, Koch D, Pionteck T, editors, Architecture of Computing Systems – ARCS 2018: 31st International Conference, Braunschweig, Germany, April 9–12, 2018, Proceedings. Braunschweig, Germany: Springer. 2018. p. 311-323. (Lecture notes in computer science). Epub 2018 Mar 8. doi: 10.1007/978-3-319-77610-1_23

A Flexible FPGA-based Inference Architecture for Pruned Deep Neural Networks

Abstract

Publication series

Conference

Access to Document

Fingerprint

Cite this