SRAM-Based Compute-in-Memory for Edge AI Processors: Trends and Design Challenges

초록

Compute-in-Memory (CIM) has been extensively investigated as a promising paradigm to overcome the intrinsic energy bottleneck of conventional von Neumann architectures. Owing to their fast access speed and strong compatibility with advanced CMOS logic processes, SRAM cells have emerged as a favorable platform for implementing SRAM-based CIM capable of performing multiply-and-accumulate (MAC) operations. In particular, analog CIM (ACIM) has attracted significant attention for edge artificial intelligence (AI) computing, primarily due to its superior energy efficiency and high inherent parallelism. Nevertheless, non-ideal effects induced by process, voltage, and temperature (PVT) variations remain critical challenges that limit computational accuracy and robustness. This paper reviews the operating principles, design challenges, and recent research trends of SRAM based ACIM technologies. © 2014 IEEE.

키워드