EUVIMEDEUVIMED entdecken
Crossref

Design of a Digital MRAM-Based Processing-in-Memory Architecture for In Situ Arithmetic and Convolution Operations

Jie Liu, Pingfan Ning, Delin Zhang, Pingjuan Niu, Yong Jiang

IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · 2026-11 · 45 (11) · 5974-5987

DOI: 10.1109/tcad.2026.3666819 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2026-11. Verfasst von Jie Liu und 4 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 47 Literaturangaben. Gefördert von: Research Program of Cangzhou Institute of Tiangong University, National Key Research and Development Program of China, National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Digital processing-in-memory (PIM) has emerged as a highly promising architecture that significantly enhances computational efficiency by embedding logic operations directly within memory. In particular, the spin-orbit torque magnetic random access memory (SOT-MRAM)-based PIM approach is regarded as a high-throughput and energy-efficient solution for data-intensive applications, effectively mitigating the inherent memory bottleneck issues in conventional computing systems. In this work, we develop low-cycle arithmetic operations using a 3T2M cell structure composed of two magnetic tunnel junctions (MTJs) that exploit both SOT and voltage-controlled magnetic anisotropy (VCMA) effects, integrated with three transistors to form a compact computational unit capable of full Boolean logic evaluation. The bitwise convolutional (CONV) operation implemented for CONV neural network (CNN) computational tasks exhibits higher parallelism and improved memory cell utilization, achieving a throughput approximately 29 times greater than that of low-cycle arithmetic operations. Furthermore, a novel comparison operation based on the 3T2M cell has been proposed and integrated into the computational module of the pooling layer within a digital MRAM-based PIM digital MRAM-PIM (DMPIM) architecture. A CNN implementation leveraging the proposed operations achieves an energy efficiency of 4.9 TOPS/W on the CIFAR-10 dataset using 16-bit precision for both inputs and weights.

Förderung

  • Research Program of Cangzhou Institute of Tiangong University
  • National Key Research and Development Program of China
  • National Natural Science Foundation of China

Zitieren

APA

Liu, J., Ning, P., Zhang, D., Niu, P., & Jiang, Y. (2026). Design of a Digital MRAM-Based Processing-in-Memory Architecture for In Situ Arithmetic and Convolution Operations. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 45(11), 5974-5987. https://doi.org/10.1109/tcad.2026.3666819

Vancouver

Liu J, Ning P, Zhang D, Niu P, Jiang Y. Design of a Digital MRAM-Based Processing-in-Memory Architecture for In Situ Arithmetic and Convolution Operations. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. 2026;45(11):5974-5987. doi:10.1109/tcad.2026.3666819

BibTeX
@article{Liu2026,
  author = {Liu, Jie and Ning, Pingfan and Zhang, Delin and Niu, Pingjuan and Jiang, Yong},
  title = {Design of a Digital MRAM-Based Processing-in-Memory Architecture for In Situ Arithmetic and Convolution Operations},
  journal = {IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems},
  year = {2026},
  volume = {45},
  number = {11},
  pages = {5974--5987},
  doi = {10.1109/tcad.2026.3666819}
}
Kostenlos starten →