MRAM-Based Analog Computing-In-Memory Macro With Reconfigurable In-Memory Reference ADC

A compact <inline-formula> <tex-math notation="LaTeX">$64\times 64$ </tex-math></inline-formula> magnetic RAM (MRAM)-based analog computing-in-memory (ACIM) macro is presented, co-optimizing in-memory techniques to overcome the inherently small read margin of MRAM-based ACIM. Based on a single word-line and dual bit-lines weight configuration, two key features are developed. The first is in-memory calibration that utilizes the array itself to mitigate sense amplifier (SA) mismatches without additional circuitry, while the proposed counter-based scheme is designed to enable efficient built-in self-calibration with minimal digital overhead. The other technique is a reconfigurable in-memory reference ADC, which generates 2-bit outputs using only a single SA by leveraging programmable reference arrays with reconfigurable levels. The macro achieves high energy efficiency, area-normalized throughput, and inference accuracy on CIFAR-10: 121.9 1b-TOPS/W, 115.1 1b-TOPS/mm2, and 84.61% for the 1b input/weight/output case, and 27.3 1b-TOPS/W, 23.2 1b-TOPS/mm2, and 90.48% for the 2b input, 3b weight, and 6b output cases, respectively. By leveraging in-memory techniques to minimize peripheral overhead, this work demonstrates that practical and scalable multi-bit ACIM can be realized with high energy and area efficiency.