New systolic arrays for matrix multiplication

Sau-Gee Chen; Jiann Cherng Lee; Chieh Chih Li

doi:10.1109/ICPP.1994.134

New systolic arrays for matrix multiplication

Sau-Gee Chen, Jiann Cherng Lee, Chieh Chih Li

National Chiao Tung University

Research output: Contribution to journal › Conference article › peer-review

2 Scopus citations

Abstract

In this paper, three new systolic arrays for matrix multiplication are proposed. The first systolic array has the minimum number of 3n-2 clock cycles in completing a matrix multiplication among the known structures, with n ² processors elements (PE's). It is achieved by applying a new input data flow and deposition scheme. The second array is derived by combining the data flow technique with the simple Blahut's matrix multiplication algorithm. Not only the second array has the least amount of processing time of 3n-2 clock cycles, it has the least area complexity of about n²/2 PE's. By further modifying its input data flow patterns, the third array is obtained. Its processing time is further reduced to 2.5n-2 clock cycles. The proposed architectures exhibit better performances than the known structures, according to several standard performance measures.

Original language	English
Article number	5727789
Journal	Proceedings of the International Conference on Parallel Processing
Volume	2
DOIs	https://doi.org/10.1109/ICPP.1994.134
State	Published - 1 Jan 1994
Event	23rd International Conference on Parallel Processing, ICPP 1994 - Raleigh, NC, United States Duration: 15 Aug 1994 → 19 Aug 1994

Access to Document

10.1109/ICPP.1994.134

Link to publication in Scopus

Fingerprint Dive into the research topics of 'New systolic arrays for matrix multiplication'. Together they form a unique fingerprint.

Cite this

@article{16ebdcdeb31241d7912c13d197480d5b,

title = "New systolic arrays for matrix multiplication",

abstract = "In this paper, three new systolic arrays for matrix multiplication are proposed. The first systolic array has the minimum number of 3n-2 clock cycles in completing a matrix multiplication among the known structures, with n 2 processors elements (PE's). It is achieved by applying a new input data flow and deposition scheme. The second array is derived by combining the data flow technique with the simple Blahut's matrix multiplication algorithm. Not only the second array has the least amount of processing time of 3n-2 clock cycles, it has the least area complexity of about n2/2 PE's. By further modifying its input data flow patterns, the third array is obtained. Its processing time is further reduced to 2.5n-2 clock cycles. The proposed architectures exhibit better performances than the known structures, according to several standard performance measures.",

author = "Sau-Gee Chen and Lee, {Jiann Cherng} and Li, {Chieh Chih}",

year = "1994",

month = jan,

day = "1",

doi = "10.1109/ICPP.1994.134",

language = "English",

volume = "2",

journal = "Proceedings of the International Conference on Parallel Processing",

issn = "0190-3918",

note = "null ; Conference date: 15-08-1994 Through 19-08-1994",

}

TY - JOUR

T1 - New systolic arrays for matrix multiplication

AU - Chen, Sau-Gee

AU - Lee, Jiann Cherng

AU - Li, Chieh Chih

PY - 1994/1/1

Y1 - 1994/1/1

N2 - In this paper, three new systolic arrays for matrix multiplication are proposed. The first systolic array has the minimum number of 3n-2 clock cycles in completing a matrix multiplication among the known structures, with n 2 processors elements (PE's). It is achieved by applying a new input data flow and deposition scheme. The second array is derived by combining the data flow technique with the simple Blahut's matrix multiplication algorithm. Not only the second array has the least amount of processing time of 3n-2 clock cycles, it has the least area complexity of about n2/2 PE's. By further modifying its input data flow patterns, the third array is obtained. Its processing time is further reduced to 2.5n-2 clock cycles. The proposed architectures exhibit better performances than the known structures, according to several standard performance measures.

AB - In this paper, three new systolic arrays for matrix multiplication are proposed. The first systolic array has the minimum number of 3n-2 clock cycles in completing a matrix multiplication among the known structures, with n 2 processors elements (PE's). It is achieved by applying a new input data flow and deposition scheme. The second array is derived by combining the data flow technique with the simple Blahut's matrix multiplication algorithm. Not only the second array has the least amount of processing time of 3n-2 clock cycles, it has the least area complexity of about n2/2 PE's. By further modifying its input data flow patterns, the third array is obtained. Its processing time is further reduced to 2.5n-2 clock cycles. The proposed architectures exhibit better performances than the known structures, according to several standard performance measures.

UR - http://www.scopus.com/inward/record.url?scp=4243593467&partnerID=8YFLogxK

U2 - 10.1109/ICPP.1994.134

DO - 10.1109/ICPP.1994.134

M3 - Conference article

AN - SCOPUS:4243593467

VL - 2

JO - Proceedings of the International Conference on Parallel Processing

JF - Proceedings of the International Conference on Parallel Processing

SN - 0190-3918

M1 - 5727789

Y2 - 15 August 1994 through 19 August 1994

ER -