DOI
https://doi.org/10.25772/PABK-JE64
Defense Date
2024
Document Type
Dissertation
Degree Name
Doctor of Philosophy
Department
Mechanical and Nuclear Engineering
First Advisor
Jayasimha Atulasimha
Abstract
Nanomagnetic devices have emerged as a promising alternative to conventional complementary metal-oxide-semiconductor (CMOS) devices due to their low energy dissipation and inherent non-volatility. However, the widespread adoption of these devices requires high-density, high-speed, reliable, scalable, and energy-efficient technologies. This thesis investigates the use of nanomagnetic memory devices as both conventional Boolean memory and multistate memory for hardware AI applications.
Magnetic tunnel junctions (MTJs) are nanomagnetic memory devices that can be switched reliably and energy-efficiently using stress-mediated switching. However, realistic material inhomogeneity and scalability pose challenges for stress-mediated switching of MTJs scaled to lateral dimensions below 50 nm. We demonstrate that resonant excitation of MTJs using surface acoustic waves can effectively address the scaling challenges associated with the high anisotropies required for sub-50 nm lateral dimensions.
For hardware AI applications, in-memory computing using non-volatile devices offers energy-efficient alternatives to traditional von Neumann computing by bridging the gap between computation and memory units. This approach addresses the significant energy inefficiency caused by the constant shuttling of data between these units. We implemented classification hardware based on deep neural networks (DNNs) using in-memory computing with domain wall (DW)-based non-volatile synaptic devices. Our research shows these DW devices can achieve multistate memories using energy efficient voltage control, however, with low resolution and stochasticity. By devising a strategy to address device stochasticity and limited precision during DNN training, we demonstrated that competitive accuracy can be achieved compared to 32-bit precision synapses, with at least two orders of magnitude energy savings. Additionally, spintronic reservoirs leveraging the rich magnetization dynamics of confined skyrmions are explored for autonomous time series prediction, offering an efficient alternative to recurrent neural networks (RNNs) and long short-term memory (LSTM) networks for edge AI applications.
Moreover, we investigate multiferroic structures and rare earth iron garnets (REIGs) for their potential in electric field control of magnetization, offering the possibility of ultra-low energy dissipation in high-density magnetic data storage applications. Our research includes the characterization of novel materials, such as bismuth-substituted yttrium iron garnet (Bi-YIG) on piezoelectric substrates, demonstrating voltage-induced control of magnetic properties. We also explore the static exchange coupling of magnetic multilayers consisting of thulium iron garnet (TmIG) with perpendicular magnetic anisotropy and cobalt iron boron (CoFeB). Incorporating REIGs into MTJs is challenging due to their insulating nature. However, coupling REIGs to a magnetic metal could enable readout of the REIG magnetization if the metal forms the free layer of an MTJ. Additionally, we explore the dynamic coupling between REIGs and CoFeB driven by spin current exchange, excited by ferromagnetic resonance, revealing that the magnetic damping parameter can be controlled by non-local relaxation and coupling.
In summary, this thesis contributes to the development of energy-efficient spintronic devices for non-volatile memory and hardware AI applications, addressing challenges in scalability, stochasticity, and material characterization, and paving the way for future advancements in these cutting-edge technologies.
Rights
© The Author
Is Part Of
VCU University Archives
Is Part Of
VCU Theses and Dissertations
Date of Submission
8-9-2024