University of North Texas
Hardware and Software Optimizations for Deep Learning Workloads on Graphics Processing Units
Abstract
dc:descriptionThis thesis investigates different hardware and software optimization techniques for accelerating deep-learning workloads using graphics processing units (GPUs). Deep learning is widely used across various engineering and scientific domains such as computer vision and natural-language processing. Such applications are used in environments spanning from resource-constrained embedded devices to exascale supercomputers. GPUs are a major hardware platform for accelerating the training and inference of deep-learning models. With the fast growth of deep-learning models, multi-GPU systems containing up to 1,000 nodes have been used for training models with billions of parameters. We characterize three different problems and techniques for improving the usage of GPUs in deep learning: deep-learning compilers, which iteratively converge to an optimized software representation of the model; a lowered L2 cache, which is aware of the duplication in image-recognition matrix-multiplication kernels; and a characterization of multi-GPU training environments across different large language models. Future research can take advantage of our results to enhance the GPU hardware and software designs and thus improve the efficiency in training and deploying deep learning models.
Degree
thesis:*- Grantor dc:publisher
- University of North Texas
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Garrigus, Justin Matthew
- Contributors dc:contributor
-
- Zhao, Hui
- Fu, Song
- Morozov, Kirill
Subjects
dc:subject × 3Rights
dc:rights- Statement dc:rights
-
- Public
- Garrigus, Justin Matthew
- Copyright
- Copyright is held by the author, unless otherwise noted. All rights Reserved.
- Language dc:language
- English
Identifiers
dc:identifier.*- Identifier
-
local-cont-no: submission_4085
https://digital.library.unt.edu/ark:/67531/metadc2415982/
ark: ark:/67531/metadc2415982