Hardware Benchmarking of Deep Neural Networks


Project Description

Artificial Intelligence is gaining popularity recently. Even though most of the efforts have been put from the software aspects, an efficient hardware execution environment is a must to realize them in practice. Two promising generic hardware solutions for efficient DNN realizations are FPGAs and GPUs. When both of them compared, FPGAs seem to provide the best efficiency/energy consumption. On the other hand, they cannot provide the enough accuracy and performance as GPUs can. Therefore, the correct hardware selection depends on the DNN specifications. At that point, benchmark results of these two architectures with respect to DNN types (e.g., detection, classification, localication, etc.) becomes very important.  Recommended Student Academic & Research Background:Basic knowledge of deep learning, programming, GPUs, FPGAs, AI, Programming​ ​​
Program - Electrical Engineering
Division - Computer, Electrical and Mathematical Sciences and Engineering
Field of Study - Electrical Engineering, Computer Science ​​

About the

Khaled Nabil Salama

Professor, Electrical and Computer Engineering<br/>Associate Dean, Computer, Electrical and Mathematical Science and Engineering

Khaled Nabil Salama
​Professor Salama's research interests cover a variety of interdisciplinary aspects of electronic circuit design and semiconductors' fabrication. He is engaged in developing devices, circuits, systems, and algorithms to enable inexpensive analytical platforms for a variety of industrial, environmental, and biomedical applications. Recently he has been working on neuromorphic circuits for brain emulation.

Desired Project Deliverables

Objectives: To provide extensive benchmarking results and comparison between FPGAs and GPUs for different DNN types. Tools to learn: Pytorch, GPUs, NVIDIA Tools, FPGAs​