Failure Mechanism and Unified Endurance Model of Embedded MRAM Towards Cache Application

Magnetic random-access memory (MRAM) shows great promise as an embedded working memory solution for data-intensive high-performance computing (HPC) and artificial intelligence (AI) applications. It requires not only high speed but also exceptional endurance performance. In this study, we discovered that current impact plays a critical role in bipolar stress (relative high voltage) and decreased while operating (relative low voltage). By integrating existing breakdown physics, a new model incorporating self-heating to estimate endurance is developed. The endurance model is used to identify the optimal operation window for maximal endurance in 1T1R MRAM cell.