Skip to main content
Video s3
    Details
    Presenter(s)
    Nanbing Pan Headshot
    Display Name
    Nanbing Pan
    Affiliation
    Affiliation
    Peking University
    Country
    Author(s)
    Display Name
    Nanbing Pan
    Affiliation
    Affiliation
    Peking University
    Display Name
    Xiaoxin Cui
    Affiliation
    Affiliation
    Peking University
    Display Name
    Xin Qiao
    Affiliation
    Affiliation
    Peking University
    Display Name
    Kanglin Xiao
    Affiliation
    Affiliation
    Peking University Shenzhen Graduate School
    Display Name
    Qingyu Guo
    Affiliation
    Affiliation
    Peking University
    Display Name
    Yuan Wang
    Affiliation
    Affiliation
    Peking University
    Abstract

    In this paper, we proposed a 28nm 64Kb SRAM based CIM macro, which supports a more complete backward propagation training algorithm. A multiply (MU) unit supports FP and BP modes. A multiply circuit (MC) supports three-inputs-multiplication (TIM) mode for the weight change analog computing. In FP and BP modes, this macro achieves an energy efficiency of 42.1TOPS/W with 2-bit input, 8-bit weight and 14-bit output multiplication and accumulation operations (MAC). In TIM mode, this macro achieves an energy efficiency of 59.4 - 2222TOPS/W with multiplication of 3 inputs and 1 output.

    Slides
    • A 28nm 64Kb SRAM Based Inference-Training Tri-Mode Computing-in-Memory Macro (application/pdf)