Year: 2020
Type: dissertation
Abstract: Deep Neural Networks (DNNs) have been proven to be state-of-the-art for many applications. DNNs are an excellent target for hardware acceleration for several reasons: they are essentially a large collection of independent multiply-accumulate operations, they require large amounts of data to be transferred, and they are deployed in an ever increasing set of applications. Hence, hardware acceleration for DNNs has been a highly active area in research and development where new designs and techniques are proposed to achieve faster and more energy-efficient DNN execution on hardware. The need to investigate such hardware acceleration techniques requires tools that allow us to experiment with such methods and to estimate their potential performance and energy benets. Accordingly, this thesis targets two main contributions: 1) DNNsim, a DNN hardware accelerators simulator to evaluate new designs, and 2) Boveda, an on-chip memory compression technique designed for Deep Learning accelerator memory hierarchies.
Source: TSpace (University of Toronto)
Author: Isak Edo Vivancos
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3152686631