Radiation Effects Modeling and Analysis in RRAM-Based In-Memory Computing Systems

RRAM-based in-memory computing architectures offer ultra-high energy efficiency for matrixvector multiplication (MVM), making them highly suitable for large-scale neural network algorithms. However, the radiation tolerance of neuromorphic computing systems has not been comprehensively assessed. This study explores the sensitivity of RRAM-based in-memory computing architectures to singleevent effects (SEE) and total ionizing dose (TID) during the inference phase. By incorporating radiation-induced conductance error models into each inference simulation, based on weight mapping rules, we analyze the impact of radiation on the inference accuracy of various neural network models. The results demonstrate that RRAM-based in-memory computing systems can effectively withstand a TID up to 30 Mrad.