A study of deep learning on multi-core processors
Ho Sy, Viet Anh
Date of Issue2016
School of Computer Engineering
The aim of this project is to conduct a study of deep learning on multi-core processors. The study is evaluated by benchmarking different deep learning frameworks on NVIDIA GPUs. The results of this project might serve as the guideline for choosing suitable deep learning frameworks to train deep learning models. Some popular deep learning frameworks and models are chosen to carry out the benchmark. For each deep learning model, the architecture is transferred into actual codes and configurations that utilize NVIDA GPUs to fasten the training process. The performance and hardware resources usage of each framework when running the models are measured and recorded in order to do the analysis and comparison later on. The results show that some frameworks outperform others in term of performance, while other frameworks demonstrate better GPU memory management. Therefore, based on the outcomes measured by this project, some frameworks should be preferred given a specific hardware details. However, this project does not benchmark the exhaustive list of all deep learning frameworks out there and should be extended to give future deep learning researchers broader view of the problem.
Final Year Project (FYP)
Nanyang Technological University