An In-Memory Computing Architecture Utilizing Energy-Efficient VGSOT MRAM Device

This brief introduces a novel 1.57-Mb IMC architecture that utilizes emerging voltage-gated spin-orbit torque magnetic random-access memory (VGSOT MRAM) device. Apart from serving as a non-volatile storage (NVM) system, this architecture facilitates various IMC operations, including logic-in-memory (LiM), ternary content addressable memory (TCAM), and in-memory dot product (IM-DP) for binary neural networks (BNN). The proposed IMC bit-cell occupies a compact area of 0.193 μ2. It achieves a write speed of 250-MHz and a read speed of 1.67-GHz for both NVM and LiM operations. The LiM functionality supports AND, NAND, OR, NOR, MAJ logical operations, and the data searches of up to 1024 bits can be performed with TCAM. Furthermore, using a BNN model with a system-level architecture of 512 (input layer) -512 (hidden layer) -10 (output layer), an impressive inference accuracy of 97.54% and 84.25% have been attained when evaluating the MNIST and FMNIST datasets, respectively. The proposed VGSOT MRAM enhances read performance and reliability, achieving a 65.74% smaller bit-cell area, 84.78%, 33.40% less read-write power, and 54.11%, 30.57% less LiM power consumption compared to 2T1R SOT-MRAM.