Skip to main content
Video s3
    Details
    Presenter(s)
    Jun-Hui Fu Headshot
    Display Name
    Jun-Hui Fu
    Affiliation
    Affiliation
    National Cheng Kung University
    Country
    Country
    Taiwan
    Author(s)
    Display Name
    Jun-Hui Fu
    Affiliation
    Affiliation
    National Cheng Kung University
    Display Name
    Soon-Jyh Chang
    Affiliation
    Affiliation
    National Cheng Kung University
    Abstract

    This paper presents a switched-capacitor charge redistribution based CIM accelerator for CNNs. The weighted capacitor switching method is proposed to reduce the number of ADCs by half, and hence diminish power consumption and area overhead of the accelerator. Another technique, called low MAC value skipping, is also implemented to skip the analog-to-digital conversion for the first few bits that further reducing the power consumption of the ADC. According to the measurement results, the energy efficiency can reach 12.02 TOPS/W with the proposed design techniques.

    Slides
    • A 12TOPS/W Computing-in-Memory Accelerator for Convolutional Neural Networks (application/pdf)