Compute-in-Memory: A Novel Computing Architecture for Future Internet-of-Things Applications

The future of the Internet-of-Things (IoT) lies in the development of intelligent and energy-efficient devices. As most IoT devices are battery-powered, achieving a balance between computing power and energy consumption is crucial. The von Neumann architecture, which serves as the general computing architecture for IoT applications, separates processing and memory units, leading to significant time and energy costs associated with data movement. This issue is further exacerbated by the recent explosive growth of data-centric applications, particularly in Artificial Intelligence (AI). Compute-In-Memory (CIM) architecture performs computational tasks directly within memory, eliminating the need for constant data shuttling, thereby significantly enhancing AI computing efficiency. This paper explores the promising applications of CIM in IoT, including enhancements in AI capabilities, revolutions in communication technologies, improvements in safety and security, and advancements in data retrieval and management. We also analyze the key challenges of CIM implementation and present a case study on image classification using the WITMEM WTM2101 CIM chip. Our experimental results demonstrate a $\mathbf{6. 8 2}$-fold improvement in energy efficiency compared to traditional von Neumann architecture, highlighting CIM’s potential in IoT applications.