Open AccessVCD-CIM: A Vertical-Compute-Decode Digital Computing-In-Memory Area-Efficient Macro for Vector-Wise Computation with Variable Accumulation Length
Author(s) -
Bo Wang,
Xiaoxue Zhong,
Jun Yang
Publication year - 2025
Publication title -
ieee access
Language(s) - English
Resource type - Magazines
SCImago Journal Rank - 0.587
H-Index - 127
eISSN - 2169-3536
DOI - 10.1109/access.2025.3574524
Subject(s) - aerospace , bioengineering , communication, networking and broadcast technologies , components, circuits, devices and systems , computing and processing , engineered materials, dielectrics and plasmas , engineering profession , fields, waves and electromagnetics , general topics for engineers , geoscience , nuclear engineering , photonics and electrooptics , power, energy and industry applications , robotics and control systems , signal processing and analysis , transportation
Computing-In-Memory (CIM) has shown significant potential in handling inference tasks for edge artificial-intelligences (Edge-AI). However, as Edge-AI tasks grow increasingly complex and diverse, with a sharp increase in edge computing requirements and limited hardware resources, there is a pressing need for CIM designs customized for Edge-AI to improve both throughput and area efficiency. Nevertheless, Edge-AI tasks involve numerous vector-wise computations with variable accumulation lengths, which obstructs enhancements in CIM area efficiency. In this paper, we introduce a solution by proposing a vertical-compute-decode digital CIM architecture (VCD-CIM) featuring vertical compute units and vertical-decode-adder-tree. This CIM architecture enables area-efficient full-precision data/vector-wise computations for Edge-AI inference tasks. Additionally, we employ a data-priority CIM strategy. In post-layout simulations, for signed 8-bit inputs, 8-bit weights, and 23-bit outputs, VCD-CIM achieves a peak energy efficiency of 30.11 TOPS/W and maintains an area energy efficiency of 1.785 TOPS/mm² across the accumulation length range of 4 to 128. We validate the effectiveness of the VCD-CIM architecture through simulations on MobileViT network, achieving CIM array activation rates between 75% and 100% during complete network inference tasks.