Author ORCID Identifier

Date Available


Year of Publication


Degree Name

Master of Science in Electrical Engineering (MSEE)

Document Type

Master's Thesis




Electrical and Computer Engineering

First Advisor

Dr. Ishan G. Thakkar


Over the past two decades, Dynamic Random-Access Memory (DRAM) has emerged as the dominant technology for implementing the main memory subsystems of all types of computing systems. However, inferring from several recent trends, computer architects in both the industry and academia have widely accepted that the density (memory capacity per chip area) and latency of DRAM based main memory subsystems cannot sufficiently scale in the future to meet the requirements of future data-centric workloads related to Artificial Intelligence (AI), Big Data, and Internet-of-Things (IoT). In fact, the achievable density and access latency in main memory subsystems presents a very fundamental trade-off. Pushing for a higher density inevitably increases access latency, and pushing for a reduced access latency often leads to a decreased density. This trade-off is so fundamental in DRAM based main memory subsystems that merely looking to re-architect DRAM subsystems cannot improve this trade-off, unless disruptive technological advancements are realized for implementing main memory subsystems.

In this thesis, we focus on two key contributions to overcome the density (represented as the total chip area for the given capacity) and access latency related challenges in main memory subsystems. First, we show that the fundamental area-latency trade-offs in DRAM can be significantly improved by redesigning the DRAM cell-array structure using the emerging monolithic 3D (M3D) integration technology. A DRAM bank structure can be split across two or more M3D-integrated tiers on the same DRAM chip, to consequently be able to significantly reduce the total on-chip area occupancy of the DRAM bank and its access peripherals. This approach is fundamentally different from the well known approach of through-silicon vias (TSVs)-based 3D stacking of DRAM tiers. This is because the M3D integration based approach does not require a separate DRAM chip per tier, whereas the 3D-stacking based approach does. Our evaluation results for PARSEC benchmarks show that our designed M3D DRAM cellarray organizations can yield up to 9.56% less latency and up to 21.21% less energy-delay product (EDP), with up to 14% less DRAM die area, compared to the conventional 2D DDR4 DRAM. Second, we demonstrate a pathway for eliminating the write disturbance errors in single-level-cell PCM, thereby positioning the PCM technology, which has inherently more relaxed density and latency trade-off compared to DRAM, as a more viable option for replacing the DRAM technology. We introduce low-temperature partial-RESET operations for writing ‘0’s in PCM cells. Compared to traditional operations that write '0's in PCM cells, partial-RESET operations do not cause disturbance errors in neighboring cells during PCM writes.

The overarching theme that connects the two individual contributions into this single thesis is the density versus latency argument. The existing PCM technology has 3 to 4× higher write latency compared to DRAM; nevertheless, the existing PCM technology can store 2 to 4 bits in a single cell compared to one bit per cell storage capacity of DRAM. Therefore, unlike DRAM, it becomes possible to increase the density of PCM without consequently increasing PCM latency. In other words, PCM exhibits inherently improved (more relaxed) density and latency trade-off. Thus, both of our contributions in this thesis, the first contribution of re-designing DRAM with M3D integration technology and the second contribution of making the PCM technology a more viable replacement of DRAM by eliminating the write disturbance errors in PCM, connect to the common overarching goal of improving the density and latency trade-off in main memory subsystems. In addition, we also discuss in this thesis possible future research directions that are aimed at extending the impacts of our proposed ideas so that they can transform the performance of main memory subsystems of the future.

Digital Object Identifier (DOI)