A Reduced Storage Row-Stationary CNN Accelerator Design With Low Wiring Complexity

碩士 === 國立中正大學 === 電機工程研究所 === 107 === Ever since its publication in 2016 ISSCC, the Row Stationary (RS) architecture has been accredited for its high performance and low energy consumption in accelerating the computation of various CNN's. As such, follow-up research works has been sustaining mo...

Full description

Bibliographic Details
Main Authors: XU, XIAO-WEN, 徐孝文
Other Authors: YEH, CHING-WEI
Format: Others
Language:zh-TW
Published: 2019
Online Access:http://ndltd.ncl.edu.tw/handle/vsnp32
id ndltd-TW-107CCU00442015
record_format oai_dc
spelling ndltd-TW-107CCU004420152019-05-16T01:31:56Z http://ndltd.ncl.edu.tw/handle/vsnp32 A Reduced Storage Row-Stationary CNN Accelerator Design With Low Wiring Complexity 考量硬體繞線複雜度之精簡儲存列駐式CNN加速設計 XU, XIAO-WEN 徐孝文 碩士 國立中正大學 電機工程研究所 107 Ever since its publication in 2016 ISSCC, the Row Stationary (RS) architecture has been accredited for its high performance and low energy consumption in accelerating the computation of various CNN's. As such, follow-up research works has been sustaining momentum. In a recent subsequent study, the Row Stationary Plus (RS+) architecture was proposed to reduce the number of registers in RS, but it incurred wire delay problem in the layout stage. In this thesis, we propose a Hierarchical Row Stationary Plus (HRS+) architecture that utilizes a two-level hierarchy to mitigate the wiring problem while preserving the register advantage of the RS+ architecture. The HRS+ is compared to redesigned RS, RS+, and a recent CNN accelerator design called COSY. Experiment results show that the HRS+ compares favorably in terms of area and power consumption. YEH, CHING-WEI 葉經緯 2019 學位論文 ; thesis 50 zh-TW
collection NDLTD
language zh-TW
format Others
sources NDLTD
description 碩士 === 國立中正大學 === 電機工程研究所 === 107 === Ever since its publication in 2016 ISSCC, the Row Stationary (RS) architecture has been accredited for its high performance and low energy consumption in accelerating the computation of various CNN's. As such, follow-up research works has been sustaining momentum. In a recent subsequent study, the Row Stationary Plus (RS+) architecture was proposed to reduce the number of registers in RS, but it incurred wire delay problem in the layout stage. In this thesis, we propose a Hierarchical Row Stationary Plus (HRS+) architecture that utilizes a two-level hierarchy to mitigate the wiring problem while preserving the register advantage of the RS+ architecture. The HRS+ is compared to redesigned RS, RS+, and a recent CNN accelerator design called COSY. Experiment results show that the HRS+ compares favorably in terms of area and power consumption.
author2 YEH, CHING-WEI
author_facet YEH, CHING-WEI
XU, XIAO-WEN
徐孝文
author XU, XIAO-WEN
徐孝文
spellingShingle XU, XIAO-WEN
徐孝文
A Reduced Storage Row-Stationary CNN Accelerator Design With Low Wiring Complexity
author_sort XU, XIAO-WEN
title A Reduced Storage Row-Stationary CNN Accelerator Design With Low Wiring Complexity
title_short A Reduced Storage Row-Stationary CNN Accelerator Design With Low Wiring Complexity
title_full A Reduced Storage Row-Stationary CNN Accelerator Design With Low Wiring Complexity
title_fullStr A Reduced Storage Row-Stationary CNN Accelerator Design With Low Wiring Complexity
title_full_unstemmed A Reduced Storage Row-Stationary CNN Accelerator Design With Low Wiring Complexity
title_sort reduced storage row-stationary cnn accelerator design with low wiring complexity
publishDate 2019
url http://ndltd.ncl.edu.tw/handle/vsnp32
work_keys_str_mv AT xuxiaowen areducedstoragerowstationarycnnacceleratordesignwithlowwiringcomplexity
AT xúxiàowén areducedstoragerowstationarycnnacceleratordesignwithlowwiringcomplexity
AT xuxiaowen kǎoliàngyìngtǐràoxiànfùzádùzhījīngjiǎnchǔcúnlièzhùshìcnnjiāsùshèjì
AT xúxiàowén kǎoliàngyìngtǐràoxiànfùzádùzhījīngjiǎnchǔcúnlièzhùshìcnnjiāsùshèjì
AT xuxiaowen reducedstoragerowstationarycnnacceleratordesignwithlowwiringcomplexity
AT xúxiàowén reducedstoragerowstationarycnnacceleratordesignwithlowwiringcomplexity
_version_ 1719177436963274752