An adaptive accuracy correction strategy in resistive random access memory (RRAM)-based computing in memory (CIM) in low-temperature scenarios

Abstract With the widespread application of artificial intelligence (AI) computing in low-temperature scenarios such as deep space and deep sea, RRAM-based edge computing has gradually attracted attention. In this paper, an adaptive reference conductance algorithm (ARCA) is proposed to improve the inference accuracy in low-temperature scenarios due to the conduction drift. The RRAM CIM chips with high read cycles are fabricated based on 28 nm CMOS logic technology, and the read times could reach 10 12 . By studying the influence of conductance drifting on inference accuracy in low temperature, a model of temperature and optimal reference conductance is proposed. Furthermore, by this model, adaptive selecting optimal reference conductance of analog-to-digital converters (ADCs) to quantize column current of RRAM array under different temperatures. At −40 ℃, the reference accuracy could increase from 75.43% to 86.8%.