Energy-Efficient Architecture for FPGA-based Deep Convolutional Neural Networks with Binary Weights

Yunzhi Duan; Shuai Li; Ruipeng Zhang; Qi Wang; Jienan Chen; Gerald E. Sobelman

doi:10.1109/ICDSP.2018.8631596

Energy-Efficient Architecture for FPGA-based Deep Convolutional Neural Networks with Binary Weights

Yunzhi Duan, Shuai Li, Ruipeng Zhang, Qi Wang, Jienan Chen, Gerald E. Sobelman

Electrical and Computer Engineering

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

Abstract

This paper presents an energy-efficient, deep parallel Convolutional Neural Network (CNN) accelerator. By adopting a recently proposed binary weight method, the CNN computations are converted into multiplication-free processing. To allow parallel accessing and storing of data, we use two RAM banks, where each bank is composed of NRAM blocks corresponding to N-parallel processing. We also design a reconfigurable CNN computing unit in a divide-and-reuse to support a variable-size convolutional filter. Compared with full-precision computing on the MNIST and CIFAR-10 classification tasks, the inference Top-1 accuracy of the binary weight CNN has dropped by 1.21% and 1.34%, respectively. The hardware implementation results show that the proposed design can achieve 2100 GOPs with a 4.6 millisecond processing latency. The deep parallel accelerator exhibits 3X energy efficiency compared to a GPU-based design.

Original language	English (US)
Title of host publication	2018 IEEE 23rd International Conference on Digital Signal Processing, DSP 2018
Publisher	Institute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)	9781538668115
DOIs	https://doi.org/10.1109/ICDSP.2018.8631596
State	Published - Jan 31 2019
Event	23rd IEEE International Conference on Digital Signal Processing, DSP 2018 - Shanghai, China Duration: Nov 19 2018 → Nov 21 2018

Publication series

Name	International Conference on Digital Signal Processing, DSP
Volume	2018-November

Conference

Conference	23rd IEEE International Conference on Digital Signal Processing, DSP 2018
Country	China
City	Shanghai
Period	11/19/18 → 11/21/18

Keywords

Convolutional Neural Network (CNN)
deep neural network
energy efficiency
parallel implementation

UN SDGs

This output contributes to the following Sustainable Development Goal(s)

Access

10.1109/ICDSP.2018.8631596

Fingerprint Dive into the research topics of 'Energy-Efficient Architecture for FPGA-based Deep Convolutional Neural Networks with Binary Weights'. Together they form a unique fingerprint.

Cite this

Duan, Y., Li, S., Zhang, R., Wang, Q., Chen, J., & Sobelman, G. E. (2019). Energy-Efficient Architecture for FPGA-based Deep Convolutional Neural Networks with Binary Weights. In 2018 IEEE 23rd International Conference on Digital Signal Processing, DSP 2018 [8631596] (International Conference on Digital Signal Processing, DSP; Vol. 2018-November). Institute of Electrical and Electronics Engineers Inc.. https://doi.org/10.1109/ICDSP.2018.8631596

Energy-Efficient Architecture for FPGA-based Deep Convolutional Neural Networks with Binary Weights. / Duan, Yunzhi; Li, Shuai; Zhang, Ruipeng; Wang, Qi; Chen, Jienan; Sobelman, Gerald E.

2018 IEEE 23rd International Conference on Digital Signal Processing, DSP 2018. Institute of Electrical and Electronics Engineers Inc., 2019. 8631596 (International Conference on Digital Signal Processing, DSP; Vol. 2018-November).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

Duan, Y, Li, S, Zhang, R, Wang, Q, Chen, J & Sobelman, GE 2019, Energy-Efficient Architecture for FPGA-based Deep Convolutional Neural Networks with Binary Weights. in 2018 IEEE 23rd International Conference on Digital Signal Processing, DSP 2018., 8631596, International Conference on Digital Signal Processing, DSP, vol. 2018-November, Institute of Electrical and Electronics Engineers Inc., 23rd IEEE International Conference on Digital Signal Processing, DSP 2018, Shanghai, China, 11/19/18. https://doi.org/10.1109/ICDSP.2018.8631596

Duan Y, Li S, Zhang R, Wang Q, Chen J, Sobelman GE. Energy-Efficient Architecture for FPGA-based Deep Convolutional Neural Networks with Binary Weights. In 2018 IEEE 23rd International Conference on Digital Signal Processing, DSP 2018. Institute of Electrical and Electronics Engineers Inc. 2019. 8631596. (International Conference on Digital Signal Processing, DSP). https://doi.org/10.1109/ICDSP.2018.8631596

Duan, Yunzhi ; Li, Shuai ; Zhang, Ruipeng ; Wang, Qi ; Chen, Jienan ; Sobelman, Gerald E. / Energy-Efficient Architecture for FPGA-based Deep Convolutional Neural Networks with Binary Weights. 2018 IEEE 23rd International Conference on Digital Signal Processing, DSP 2018. Institute of Electrical and Electronics Engineers Inc., 2019. (International Conference on Digital Signal Processing, DSP).

@inproceedings{13aa6ca9ed0c4a2fb4dcec4e8ad86ece,

title = "Energy-Efficient Architecture for FPGA-based Deep Convolutional Neural Networks with Binary Weights",

abstract = "This paper presents an energy-efficient, deep parallel Convolutional Neural Network (CNN) accelerator. By adopting a recently proposed binary weight method, the CNN computations are converted into multiplication-free processing. To allow parallel accessing and storing of data, we use two RAM banks, where each bank is composed of NRAM blocks corresponding to N-parallel processing. We also design a reconfigurable CNN computing unit in a divide-and-reuse to support a variable-size convolutional filter. Compared with full-precision computing on the MNIST and CIFAR-10 classification tasks, the inference Top-1 accuracy of the binary weight CNN has dropped by 1.21% and 1.34%, respectively. The hardware implementation results show that the proposed design can achieve 2100 GOPs with a 4.6 millisecond processing latency. The deep parallel accelerator exhibits 3X energy efficiency compared to a GPU-based design.",

keywords = "Convolutional Neural Network (CNN), deep neural network, energy efficiency, parallel implementation",

author = "Yunzhi Duan and Shuai Li and Ruipeng Zhang and Qi Wang and Jienan Chen and Sobelman, {Gerald E.}",

year = "2019",

month = jan,

day = "31",

doi = "10.1109/ICDSP.2018.8631596",

language = "English (US)",

series = "International Conference on Digital Signal Processing, DSP",

publisher = "Institute of Electrical and Electronics Engineers Inc.",

booktitle = "2018 IEEE 23rd International Conference on Digital Signal Processing, DSP 2018",

note = "23rd IEEE International Conference on Digital Signal Processing, DSP 2018 ; Conference date: 19-11-2018 Through 21-11-2018",

}

TY - GEN

T1 - Energy-Efficient Architecture for FPGA-based Deep Convolutional Neural Networks with Binary Weights

AU - Duan, Yunzhi

AU - Li, Shuai

AU - Zhang, Ruipeng

AU - Wang, Qi

AU - Chen, Jienan

AU - Sobelman, Gerald E.

PY - 2019/1/31

Y1 - 2019/1/31

N2 - This paper presents an energy-efficient, deep parallel Convolutional Neural Network (CNN) accelerator. By adopting a recently proposed binary weight method, the CNN computations are converted into multiplication-free processing. To allow parallel accessing and storing of data, we use two RAM banks, where each bank is composed of NRAM blocks corresponding to N-parallel processing. We also design a reconfigurable CNN computing unit in a divide-and-reuse to support a variable-size convolutional filter. Compared with full-precision computing on the MNIST and CIFAR-10 classification tasks, the inference Top-1 accuracy of the binary weight CNN has dropped by 1.21% and 1.34%, respectively. The hardware implementation results show that the proposed design can achieve 2100 GOPs with a 4.6 millisecond processing latency. The deep parallel accelerator exhibits 3X energy efficiency compared to a GPU-based design.

AB - This paper presents an energy-efficient, deep parallel Convolutional Neural Network (CNN) accelerator. By adopting a recently proposed binary weight method, the CNN computations are converted into multiplication-free processing. To allow parallel accessing and storing of data, we use two RAM banks, where each bank is composed of NRAM blocks corresponding to N-parallel processing. We also design a reconfigurable CNN computing unit in a divide-and-reuse to support a variable-size convolutional filter. Compared with full-precision computing on the MNIST and CIFAR-10 classification tasks, the inference Top-1 accuracy of the binary weight CNN has dropped by 1.21% and 1.34%, respectively. The hardware implementation results show that the proposed design can achieve 2100 GOPs with a 4.6 millisecond processing latency. The deep parallel accelerator exhibits 3X energy efficiency compared to a GPU-based design.

KW - Convolutional Neural Network (CNN)

KW - deep neural network

KW - energy efficiency

KW - parallel implementation

UR - http://www.scopus.com/inward/record.url?scp=85062797447&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85062797447&partnerID=8YFLogxK

U2 - 10.1109/ICDSP.2018.8631596

DO - 10.1109/ICDSP.2018.8631596

M3 - Conference contribution

AN - SCOPUS:85062797447

T3 - International Conference on Digital Signal Processing, DSP

BT - 2018 IEEE 23rd International Conference on Digital Signal Processing, DSP 2018

PB - Institute of Electrical and Electronics Engineers Inc.

T2 - 23rd IEEE International Conference on Digital Signal Processing, DSP 2018

Y2 - 19 November 2018 through 21 November 2018

ER -