Towards Efficient SOT-Assisted STT-MRAM Cell Switching Using Reinforcement Learning

Nonvolatile memory is a promising candidate to replace CMOS devices with the two most common magnetoresistive RAM (MRAM) cell types being spin-transfer torque (STT-)MRAM and spin-orbit torque (SOT-)MRAM. Recently introduced combinations of these two mechanisms allow the amelioration of MRAM performance. To optimize the switching, we implemented an approach based on reinforcement learning. In particular, we train an agent to switch an SOT-assisted STT-MRAM cell by applying STT or SOT current pulses independently. During this process, by means of rewards based on its actions, the agent is encouraged to either reverse the magnetization in the memory cell fast or by using little energy. After successfully training RL agents under the given constraints, the results of the two rewarding schemes applied to SOT-assisted STT-MRAM are discussed.