Dissertation
HIGH PERFORMANCE AND RELIABLE PROCESSING-IN-MEMORY ACCELERATORS FOR GRAPH-BASED MACHINE LEARNING
Washington State University
Doctor of Philosophy (PhD), Washington State University
05/2025
DOI:
https://doi.org/10.7273/000007443
Abstract
Graph-based machine learning has emerged as a critical tool for solving complex problems in domains such as social networks, recommendation systems, and biological research. These applications require the processing of massive, irregularly structured datasets, posing significant challenges for traditional von Neumann architectures due to frequent memory access and data movement bottlenecks. Processing-in-Memory (PIM) accelerators offer a promising solution by enabling computation directly within memory, thereby reducing data movement and improving efficiency. However, achieving both high performance and reliability in PIM designs is crucial to meet the demands of graph-based workloads, which often involve diverse and dynamic computations. This dissertation explores the design of high-performance and reliable PIM accelerators tailored for graph-based machine learning, addressing challenges such as irregular data access patterns, fault tolerance, and scalability, while pushing the boundaries of energy efficiency and computational throughput. The first part of this dissertation focuses on the use of model and data pruning for enabling energy-efficient and high-performance acceleration of Graph Neural Networks on ReRAM-based PIM architectures.
Later in this dissertation, we discuss the challenges associated with Neural Network training and inference using existing PIM-based architectures, as well as the inherent reliability and non-ideal behavior of PIM devices. Finally, we present a heterogeneous PIM-based architecture that combines more than one type of PIM device and achieves a balanced tradeoff between performance, power, area, and predictive accuracy compared to homogeneous counterparts.
Overall, in this dissertation we demonstrate various methods enable the design of PIM-based manycore architectures optimized for high-performance machine learning and big data workloads.
Metrics
6 File views/ downloads
16 Record Views
Details
- Title
- HIGH PERFORMANCE AND RELIABLE PROCESSING-IN-MEMORY ACCELERATORS FOR GRAPH-BASED MACHINE LEARNING
- Creators
- Chukwufumnanya osazee Ogbogu
- Contributors
- Partha Pratim Pande (Chair)Janardhan Rao Doppa (Co-Chair)Dae Hyun Kim (Committee Member)Biresh Kumar Joardar (Committee Member)
- Awarding Institution
- Washington State University
- Academic Unit
- School of Electrical Engineering and Computer Science
- Theses and Dissertations
- Doctor of Philosophy (PhD), Washington State University
- Publisher
- Washington State University
- Number of pages
- 174
- Identifiers
- 99901221254001842
- Language
- English
- Resource Type
- Dissertation