Edge-Optimized AI Architecture: MRAM-based Near Memory Computing Core Balancing between Memory Capacity and Computation

General-purpose edge neural networks require a lightweight architecture, delivering both suitable memory capacity and compute resources. Typically, the SRAM-based computing-in memory (CIM) architecture is difficult to provide sufficient storage to fulfilling the computing requirements, and various buffers are needed. In this work, we present a novel high-capacity MRAM-based near memory computing (NMC) architecture with configurable computing unit, adeptly striking a balance between the edge network’s demands for storage and computing. It has realized a storage density of $1.49 \mathrm{Mb} / \mathrm{mm}^{2}$ on-chip and an excellent energy efficiency of 6.164 TOPS/W.