Comparative Study on Analog and Digital Processing-in-Memory Architectures

Abstract

Processing-in-Memory (PIM) is a promising solution to the memory bottleneck problem in modern computing systems. DRAM-based PIM technology can be implemented in two ways: analog and digital. We discuss the advantages and challenges of each PIM architecture, and provide a comprehensive evaluation of the two PIM architectures, comparing their performance across various types of operations, including bitwise and arithmetic operations. Our evaluation results reveal that analog PIM has higher throughput than digital PIM for bitwise operations. However, for arithmetic operations, digital PIM shows higher throughput than analog PIM in HBM2 for all ADD and MUL operations across various bit-width configurations. The choice of using analog or digital PIM should depend on the target operation, operand bit-width, and the specifications of the DRAM device. This study provides new insights into PIM technology to facilitate future research in this field.