Shuffle-CIM: An All-Analog Computing-in-Memory Macro With Shuffle Scheme and Analog Non-MAC Implementation for Smart Sensing

Erxiang Ren, Daniel Zheng Fang, Yiming Liu, Xinzhe Wang, Tong Wang, Yiexian Tay, Hanwen Li, Li Luo, Hongwei Guo, Ying Fu, Qi Wei, Ge Shi, Fei Qiao

Abstract: Smart sensing is an emerging technology that enables real-time intelligent decision-making directly at the data source. However, smart sensors typically face difficulties in the deployment of complex algorithms, such as large convolutional neural networks (CNNs), owing to a restricted power and area budget. Although analog computing-in-memory (CIM) has made great strides in these areas, a major remaining cost of energy and area stems from partial sums (pSUMs) and their associated digital circuitry. This brief presents Shuffle-CIM, a CIM architecture that incorporates the shuffle scheme and analog non-multiply-and-accumulate (non-MAC) implementation. The shuffle scheme eliminates the need for pSUM accumulation, and without pSUM constraints, a specialized ReLU-ADC design integrates non-MAC computations directly into the analog domain, achieving an all-analog CIM macro. Implemented on TSMC 40nm, the Shuffle-CIM design achieves a power consumption of 4.90mW and an area of $0.396\mathrm {mm^{2}}$ , representing a $3.1\times $ and $2.9\times $ reduction compared to the conventional CIM architecture. The resulting energy efficiency and area efficiency reach $107.55\mathrm {TOPS/W}$ and $1.33 \mathrm {TOPS/mm^{2}}$ , with an accuracy of 89.01% on CIFAR-10.

External IDs:doi:10.1109/tcsii.2025.3607321