DM-IMCA: A Dual-mode In-Memory Computing Architecture for General Purpose Processing

Abstract

In-Memory Computing (IMC) architectures based on Static Random Access Memory (SRAM) can improve system performance and energy-efficiency dramatically. However, most of the existing SRAM-based implementations are designed for specific purposes like accelerating neural networks, which limits the application scenarios of IMC. In this paper, we propose DM-IMCA, a novel IMC architecture with two work modes for general purpose processing. It utilizes our proposed 9T bitcell based computational SRAM as the location to perform IMC operations. Besides, a new IMC Instruction Set Architecture (ISA) as well as an automated vector computing mechanism are also proposed to facilitate DM-IMCA's programming and accelerate in-memory computing, respectively. The simulation results show that DM-IMCA can bring a performance increase by up to 257x, and SRAM energy saving by up to 3x, compared to a baseline system.