A Dynamic Power-Only Compute-in-Memory Macro With Power-of-Two Nonlinear SAR ADC for Nonvolatile Ferroelectric Capacitive Crossbar Array
Injune Yeo, Wangxin He, Yuan-Chun Luo, Shimeng Yu, Jae-sun Seo
Abstract
Analog computing-in-memory (CIM) using emerging resistive nonvolatile memory (NVM) technologies faces challenges such as static power consumption, current flow-induced IR drop, and the need for multiple power-hungry ADCs. In this letter, we present ferroelectric capacitive array (FCA)-based energy/area efficient CIM macro used for charge-domain multiply-and-accumulate operations, which addresses the challenges of resistive NVM CIMs. The proposed CIM macro involves encoding ternary input activations and weights into voltages, and enabling parasitic insensitive charge readout. A power-of-two nonlinear SAR ADC is introduced, designed for energy-efficiency and hardware-friendliness. This ADC employs adaptive conversion skipping based on input voltage, resulting in fine precision for concentrated input levels and coarse conversion for sparse input levels. The proposed FCA-based CIM macro in 180nm CMOS demonstrates 16×8 analog MAC operation with an energy-efficiency of 1.75 TOPS/W and classification accuracy of 90.2% is obtained for CIFAR-10 dataset.