High-Density STT-Assisted SOT-MRAM (SAS-MRAM) for Energy-Efficient AI Applications

Energy-efficient computing is essential for addressing the rising power demands of modern data-intensive applications and ensuring sustainable technology advancement. Magnetoresistive random access memory (MRAM) has emerged as a pivotal technology in this domain, offering nonvolatile memory solutions that combine low power consumption with high performance. Spin-orbit torque (SOT) MRAM (SOT-MRAM) and its variants stand out for its potential to deliver SRAM-like performance at a higher bit-cell density. In this article, we present a novel high-density STT-assisted SOT-MRAM (SAS-MRAM) technology designed for energy-efficient artificial intelligence (AI) applications. SAS-MRAM capitalizes on the advantages of both spin-transfer torque (STT) and SOT mechanisms, utilizing a multi-bit-shared SOT line to achieve high-speed, high-density, and high-endurance memory performance. Our experimental results validate the potential of SAS-MRAM to address the limitations of current memory technologies. An AI application of ResNet-18 deployed in SAS-MRAM shows <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\sim 32.7\times $ </tex-math></inline-formula> energy-delay-product (EDP) benefits compared to that in SRAM, presenting a promising solution for future AI hardware implementations, especially at edge where low-power training and inference of AI models are necessary.