Bit-width reduction and customized register for low cost convolutional neural network accelerator

Kyungrak Choi, Woong Choi, Kyungho Shin, Jongsun Park

Research output: Chapter in Book/Report/Conference proceedingConference contribution

2 Citations (Scopus)

Abstract

This paper presents a low area and energy efficient hardware accelerator for the deep convolutional neural networks (CNNs). Based on the multiply-accumulate (MAC) based architecture, three design techniques are proposed to reduce the hardware cost of the convolutional computations. First, to reduce the computational bit-width of convolutions, an adaptive bit-width reduction scheme is proposed based on differential input method. The bit-width reduction approach can reduce the 37 % of operation bit-width with almost ignorable CNN accuracy degradation. Second, it has been found that adapting bi-directional filtering window in CNN accelerator can considerably reduce the energy for data movement with much smaller number of memory accesses. To expedite the bi-directional filtering operations, we also propose a bidirectional first-input-first-output (bi-FIFO). With SRAM bit-cell layout manner, the proposed bi-FIFO facilitates fast data re-distribution with area and energy efficiency. To verify the effectiveness of the proposed techniques, the AlexNet accelerator has been designed. The numerical results show that the proposed adaptive bit-width reduction scheme achieves 25.9% and 47.3% of area and energy savings, respectively. The bi-FIFO based accelerator also achieves 33 % improved processing time.

Original languageEnglish
Title of host publicationISLPED 2017 - IEEE/ACM International Symposium on Low Power Electronics and Design
PublisherInstitute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)9781509060238
DOIs
Publication statusPublished - 2017 Aug 11
Event22nd IEEE/ACM International Symposium on Low Power Electronics and Design, ISLPED 2017 - Taipei, Taiwan, Province of China
Duration: 2017 Jul 242017 Jul 26

Other

Other22nd IEEE/ACM International Symposium on Low Power Electronics and Design, ISLPED 2017
CountryTaiwan, Province of China
CityTaipei
Period17/7/2417/7/26

Fingerprint

Particle accelerators
Neural networks
Costs
Hardware
Static random access storage
Convolution
Energy efficiency
Energy conservation
Data storage equipment
Degradation
Processing

Keywords

  • Convolutional Neural Network
  • Deep Neural Network
  • Energy Efficiency
  • FIFO
  • Filter
  • Line Buffer
  • Weight

ASJC Scopus subject areas

  • Engineering(all)

Cite this

Choi, K., Choi, W., Shin, K., & Park, J. (2017). Bit-width reduction and customized register for low cost convolutional neural network accelerator. In ISLPED 2017 - IEEE/ACM International Symposium on Low Power Electronics and Design [8009164] Institute of Electrical and Electronics Engineers Inc.. https://doi.org/10.1109/ISLPED.2017.8009164

Bit-width reduction and customized register for low cost convolutional neural network accelerator. / Choi, Kyungrak; Choi, Woong; Shin, Kyungho; Park, Jongsun.

ISLPED 2017 - IEEE/ACM International Symposium on Low Power Electronics and Design. Institute of Electrical and Electronics Engineers Inc., 2017. 8009164.

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Choi, K, Choi, W, Shin, K & Park, J 2017, Bit-width reduction and customized register for low cost convolutional neural network accelerator. in ISLPED 2017 - IEEE/ACM International Symposium on Low Power Electronics and Design., 8009164, Institute of Electrical and Electronics Engineers Inc., 22nd IEEE/ACM International Symposium on Low Power Electronics and Design, ISLPED 2017, Taipei, Taiwan, Province of China, 17/7/24. https://doi.org/10.1109/ISLPED.2017.8009164
Choi K, Choi W, Shin K, Park J. Bit-width reduction and customized register for low cost convolutional neural network accelerator. In ISLPED 2017 - IEEE/ACM International Symposium on Low Power Electronics and Design. Institute of Electrical and Electronics Engineers Inc. 2017. 8009164 https://doi.org/10.1109/ISLPED.2017.8009164
Choi, Kyungrak ; Choi, Woong ; Shin, Kyungho ; Park, Jongsun. / Bit-width reduction and customized register for low cost convolutional neural network accelerator. ISLPED 2017 - IEEE/ACM International Symposium on Low Power Electronics and Design. Institute of Electrical and Electronics Engineers Inc., 2017.
@inproceedings{566f3fcd630c4c2c9662c8f5d38eea4e,
title = "Bit-width reduction and customized register for low cost convolutional neural network accelerator",
abstract = "This paper presents a low area and energy efficient hardware accelerator for the deep convolutional neural networks (CNNs). Based on the multiply-accumulate (MAC) based architecture, three design techniques are proposed to reduce the hardware cost of the convolutional computations. First, to reduce the computational bit-width of convolutions, an adaptive bit-width reduction scheme is proposed based on differential input method. The bit-width reduction approach can reduce the 37 {\%} of operation bit-width with almost ignorable CNN accuracy degradation. Second, it has been found that adapting bi-directional filtering window in CNN accelerator can considerably reduce the energy for data movement with much smaller number of memory accesses. To expedite the bi-directional filtering operations, we also propose a bidirectional first-input-first-output (bi-FIFO). With SRAM bit-cell layout manner, the proposed bi-FIFO facilitates fast data re-distribution with area and energy efficiency. To verify the effectiveness of the proposed techniques, the AlexNet accelerator has been designed. The numerical results show that the proposed adaptive bit-width reduction scheme achieves 25.9{\%} and 47.3{\%} of area and energy savings, respectively. The bi-FIFO based accelerator also achieves 33 {\%} improved processing time.",
keywords = "Convolutional Neural Network, Deep Neural Network, Energy Efficiency, FIFO, Filter, Line Buffer, Weight",
author = "Kyungrak Choi and Woong Choi and Kyungho Shin and Jongsun Park",
year = "2017",
month = "8",
day = "11",
doi = "10.1109/ISLPED.2017.8009164",
language = "English",
booktitle = "ISLPED 2017 - IEEE/ACM International Symposium on Low Power Electronics and Design",
publisher = "Institute of Electrical and Electronics Engineers Inc.",

}

TY - GEN

T1 - Bit-width reduction and customized register for low cost convolutional neural network accelerator

AU - Choi, Kyungrak

AU - Choi, Woong

AU - Shin, Kyungho

AU - Park, Jongsun

PY - 2017/8/11

Y1 - 2017/8/11

N2 - This paper presents a low area and energy efficient hardware accelerator for the deep convolutional neural networks (CNNs). Based on the multiply-accumulate (MAC) based architecture, three design techniques are proposed to reduce the hardware cost of the convolutional computations. First, to reduce the computational bit-width of convolutions, an adaptive bit-width reduction scheme is proposed based on differential input method. The bit-width reduction approach can reduce the 37 % of operation bit-width with almost ignorable CNN accuracy degradation. Second, it has been found that adapting bi-directional filtering window in CNN accelerator can considerably reduce the energy for data movement with much smaller number of memory accesses. To expedite the bi-directional filtering operations, we also propose a bidirectional first-input-first-output (bi-FIFO). With SRAM bit-cell layout manner, the proposed bi-FIFO facilitates fast data re-distribution with area and energy efficiency. To verify the effectiveness of the proposed techniques, the AlexNet accelerator has been designed. The numerical results show that the proposed adaptive bit-width reduction scheme achieves 25.9% and 47.3% of area and energy savings, respectively. The bi-FIFO based accelerator also achieves 33 % improved processing time.

AB - This paper presents a low area and energy efficient hardware accelerator for the deep convolutional neural networks (CNNs). Based on the multiply-accumulate (MAC) based architecture, three design techniques are proposed to reduce the hardware cost of the convolutional computations. First, to reduce the computational bit-width of convolutions, an adaptive bit-width reduction scheme is proposed based on differential input method. The bit-width reduction approach can reduce the 37 % of operation bit-width with almost ignorable CNN accuracy degradation. Second, it has been found that adapting bi-directional filtering window in CNN accelerator can considerably reduce the energy for data movement with much smaller number of memory accesses. To expedite the bi-directional filtering operations, we also propose a bidirectional first-input-first-output (bi-FIFO). With SRAM bit-cell layout manner, the proposed bi-FIFO facilitates fast data re-distribution with area and energy efficiency. To verify the effectiveness of the proposed techniques, the AlexNet accelerator has been designed. The numerical results show that the proposed adaptive bit-width reduction scheme achieves 25.9% and 47.3% of area and energy savings, respectively. The bi-FIFO based accelerator also achieves 33 % improved processing time.

KW - Convolutional Neural Network

KW - Deep Neural Network

KW - Energy Efficiency

KW - FIFO

KW - Filter

KW - Line Buffer

KW - Weight

UR - http://www.scopus.com/inward/record.url?scp=85028582850&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85028582850&partnerID=8YFLogxK

U2 - 10.1109/ISLPED.2017.8009164

DO - 10.1109/ISLPED.2017.8009164

M3 - Conference contribution

BT - ISLPED 2017 - IEEE/ACM International Symposium on Low Power Electronics and Design

PB - Institute of Electrical and Electronics Engineers Inc.

ER -