Spin based DNN accelerator with bit splitting technique
Physica Scripta · 19.06.2026 · 101 (24) · 246005
DOI: 10.1088/1402-4896/ae7511 ↗Kurz gefasst
Fachartikel in der Zeitschrift Physica Scripta, online erschienen am 16.06.2026. Verfasst von Gaurav Verma und 2 weiteren Autor:innen. Die Arbeit nennt 35 Literaturangaben.
Abstract (Originalsprache)
Spin crossbar arrays have gained significant attention for accelerating deep neural networks (DNNs) in neuromorphic hardware due to their ability to perform in-memory computation. However, practical DNN implementations require multi-bit precision, which, when mapped onto crossbar architectures, leads to increased peripheral complexity, higher energy consumption due to ADC/DACs, and limited scalability. To address these challenges, we propose a multi-level spin–orbit torque magnetic random-access memory (SOT-MRAM)-based computing-in-memory (CiM) architecture that uses bit-splitting to map multi-bit weights into parallel single-bit operations, enabling efficient parallel processing and improved energy efficiency. Unlike conventional CiM approaches based on binary or resistive memories, this work demonstrates a dual-bit-per-cell SOT-MRAM architecture for crossbar-level computation, improving area and power efficiency. The design optimizes four-level conductance states and incorporates input-splitting-based peripheral schemes for reliable multi-level operation. Through algorithm–hardware co-design, the architecture achieves high classification accuracy ∼89.5% with reduced area and power consumption, demonstrating up to 3× improvement in energy efficiency over state-of-the-art CiM implementations and highlighting the potential of MLC SOT-MRAM with bit-splitting for scalable, energy-efficient AI hardware.
Zitieren
APA
Verma, G., Soni, S., & Kaushik, B. K. (2026). Spin based DNN accelerator with bit splitting technique. Physica Scripta, 101(24), 246005. https://doi.org/10.1088/1402-4896/ae7511
Vancouver
Verma G, Soni S, Kaushik BK. Spin based DNN accelerator with bit splitting technique. Physica Scripta. 2026;101(24):246005. doi:10.1088/1402-4896/ae7511
BibTeX
@article{Verma2026,
author = {Verma, Gaurav and Soni, Sandeep and Kaushik, Brajesh Kumar},
title = {Spin based DNN accelerator with bit splitting technique},
journal = {Physica Scripta},
year = {2026},
volume = {101},
number = {24},
pages = {246005},
doi = {10.1088/1402-4896/ae7511}
}
EUVIMED