Skip to main navigation Skip to search Skip to main content

A High Accuracy Low Power Convolution Operator with 12T SRAM for CNN

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

To get high accuracy, various weight should be stored in memory. For this, this paper presents tri-state weight static random access memory(SRAM). 12T SRAM is a form of power gating on the conventional 10T SRAM. By using power gating, the inverter can be turned off. The new weight (0) can be stored in 12T SRAM when inverter is turned off. The operator fabricated in a 0.18-μm CMOS process dissipates 172.3μW with the supply of 1.8V while convolution. Even without sizing, the writing margin is better than the conventional SRAM and the accuracy is improved by 23.2%.

Original languageEnglish
Title of host publicationICUFN 2021 - 2021 12th International Conference on Ubiquitous and Future Networks
PublisherIEEE Computer Society
Pages295-298
Number of pages4
ISBN (Electronic)9781728164762
DOIs
StatePublished - 17 Aug 2021
Event12th International Conference on Ubiquitous and Future Networks, ICUFN 2021 - Virtual, Jeju Island, Korea, Republic of
Duration: 17 Aug 202120 Aug 2021

Publication series

NameInternational Conference on Ubiquitous and Future Networks, ICUFN
Volume2021-August
ISSN (Print)2165-8528
ISSN (Electronic)2165-8536

Conference

Conference12th International Conference on Ubiquitous and Future Networks, ICUFN 2021
Country/TerritoryKorea, Republic of
CityVirtual, Jeju Island
Period17/08/2120/08/21

Keywords

  • 12T SRAM
  • Convolution
  • convolutional neural networks (CNN)
  • processing in memory (PIM)
  • Tri-state weight

Fingerprint

Dive into the research topics of 'A High Accuracy Low Power Convolution Operator with 12T SRAM for CNN'. Together they form a unique fingerprint.

Cite this