Presentation 2015-06-20
A Deep Convolutional Neural Network Based on Nested Residue Number System
Hiroki Nakahara, Tsutomu Sasao,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) A pre-trained deep convolutional neural network~(DCNN) is the feedforward computation perspective which is widely used for the embedded systems. In the DCNN, a 2D convolutional operation occupies more than 90% of the computation time. Since the 2D convolutional operation consumes many multiply-accumulation~(MAC) units, conventional realizations could not realize a fully parallel DCNN. In this paper, we propose the nested residue number system~(nested RNS). It is a new type of RNS which decomposes the MAC units. In this paper, 48bit MAC units are decomposed into parallel 4bit ones realized by look-up tables on the FPGA. Also, we show the binary to nested RNS converter realized by on-chip BRAMs, while the nested RNS to binary one realized by DSP blocks and BRAMs. Since our architecture uses most of the FPGA resources, the resource utilization efficiency is very high. We implemented the ImageNet DCNN using the nested RNS on a Xilinx Virtex VC707 evaluation board. As for the performance per area measure~(GOPS~(Giga operations per second) per a slice), the proposed one is 5.81 times better than the existing best realization.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) FPGA / Convolutional Neural Netowrk / Deep Neural Network / Residue Number System / Nested RNS
Paper # RECONF2015-17
Date of Issue 2015-06-12 (RECONF)

Conference Information
Committee RECONF
Conference Date 2015/6/19(2days)
Place (in Japanese) (See Japanese page)
Place (in English) Kyoto University
Topics (in Japanese) (See Japanese page)
Topics (in English) the 10th anniversary celebration of RECONF: Reconfigurable Systems, etc.
Chair Minoru Watanabe(Shizuoka Univ.)
Vice Chair Masato Motomura(Hokkaido Univ.) / Yuichiro Shibata(Nagasaki Univ.)
Secretary Masato Motomura(Toshiba) / Yuichiro Shibata(Univ. of Tsukuba)
Assistant Kazuya Tanikagawa(Hiroshima City Univ.) / Takefumi Miyoshi(e-trees.Japan)

Paper Information
Registration To Technical Committee on Reconfigurable Systems
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) A Deep Convolutional Neural Network Based on Nested Residue Number System
Sub Title (in English)
Keyword(1) FPGA
Keyword(2) Convolutional Neural Netowrk
Keyword(3) Deep Neural Network
Keyword(4) Residue Number System
Keyword(5) Nested RNS
1st Author's Name Hiroki Nakahara
1st Author's Affiliation Ehime University(Ehime Univ.)
2nd Author's Name Tsutomu Sasao
2nd Author's Affiliation Meiji Univeristy(Meiji Univ.)
Date 2015-06-20
Paper # RECONF2015-17
Volume (vol) vol.115
Number (no) RECONF-109
Page pp.pp.91-96(RECONF),
#Pages 6
Date of Issue 2015-06-12 (RECONF)