DeepNVM++: Cross-Layer Modeling and Optimization Framework of Nonvolatile Memories for Deep Learning

Inci, Ahmet; Isgenc, Mehmet Meric; Marculescu, Diana

doi:10.1109/tcad.2021.3127148

Cited by 8 publications

(1 citation statement)

References 62 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Therefore, there is a need to perform a sensitivity analysis on the parameters that vary from DRAM to SOT-MRAM. In this study, a conservative scaling of SOT-MRAM parameters from DRAM-1600 is adopted using the methodology in [1], [34]. The scaling methodology was validated in the work [1] and [23].…”

Section: A Parameters For Estimation Of Timingmentioning

confidence: 99%

A Detailed Study of SOT-MRAM as an Alternative to DRAM Primary Memory in Multi-Core Environment

Kallinatha,

Rai,

Talawar

2024

IEEE Access

View full text Add to dashboard Cite

Alternative memory technologies must be explored to foster the need for modern applications. This work proposes a novel approach incorporating Spin-Orbit-Torque-Magnetic-RAM (SOT-MRAM) into hybrid and full main memory architectures within a multi-core system, encompassing various memory configurations and capacities. The study addresses the challenge of evaluating SOT-MRAM-based memory systems when specific SOT-MRAM memory parameters are not publicly available. The research methodology encompasses micro-architectural (circuit-level) design space exploration and comprehensive full system simulations, which evaluate benchmark programs representing diverse application domains. The evaluation includes three memory structures with varying memory organizations and capacities. The results underscore SOT-MRAM as a robust replacement for DRAM or hybrid memory, offering compelling advantages such as a remarkable 74.05% reduction in power consumption, a noteworthy 40.10% increase in bandwidth utilization, and a significant 72.85% reduction in Energy-Delay Product (EDP). Additionally, the incurred maximum latency penalties are minimal, with a 3.71% increase for hybrid structures and a mere 0.07% for standalone SOT-MRAM memory structures.

show abstract