Low Cost Early Exit Decision Unit Design for CNN Accelerator

Geonho Kim, Jongsun Park

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

    9 Citations (Scopus)

    Abstract

    Early exit has been studied as a way to reduce the complex computation of convolutional neural networks. However, in order to determine whether to exit early in a conventional CNN accelerator, there is a problem that a unit for computing softmax layer having a large hardware overhead is required. To solve this problem, we propose a low cost early exit decision unit. The proposed architecture uses only fully-connected (FC) layer outputs to make early exit decisions, so the computation of the softmax layer is not necessary. Our implementation results show an energy reduction of 68% with an accuracy drop of less than 0.3%.

    Original languageEnglish
    Title of host publicationProceedings - International SoC Design Conference, ISOCC 2020
    PublisherInstitute of Electrical and Electronics Engineers Inc.
    Pages127-128
    Number of pages2
    ISBN (Electronic)9781728183312
    DOIs
    Publication statusPublished - 2020 Oct 21
    Event17th International System-on-Chip Design Conference, ISOCC 2020 - Yeosu, Korea, Republic of
    Duration: 2020 Oct 212020 Oct 24

    Publication series

    NameProceedings - International SoC Design Conference, ISOCC 2020

    Conference

    Conference17th International System-on-Chip Design Conference, ISOCC 2020
    Country/TerritoryKorea, Republic of
    CityYeosu
    Period20/10/2120/10/24

    Bibliographical note

    Publisher Copyright:
    © 2020 IEEE.

    Keywords

    • CNN accelerator
    • early exit
    • softmax

    ASJC Scopus subject areas

    • Energy Engineering and Power Technology
    • Electrical and Electronic Engineering
    • Instrumentation
    • Artificial Intelligence
    • Hardware and Architecture

    Fingerprint

    Dive into the research topics of 'Low Cost Early Exit Decision Unit Design for CNN Accelerator'. Together they form a unique fingerprint.

    Cite this