Politecnico di Torino (logo)

Modeling and Mapping of a GoogLeNet CNN on a Grid of Processing Cells

Claudio Raccomandato

Modeling and Mapping of a GoogLeNet CNN on a Grid of Processing Cells.

Rel. Guido Masera, Rainer Dömer. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Elettronica (Electronic Engineering), 2022

PDF (Tesi_di_laurea) - Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives.

Download (3MB) | Preview

System-level design methodologies evolve in response to increasing complexity of applications. Transaction-level modeling (TLM) is one technique that allows the designer to capture the specifications of complex digital systems without defining low-level implementation details. The Grid of Processing Cells (GPC) has been proposed as a highly scalable many-core architecture and is modeled using SystemC TLM-2.0 methodology. This thesis describes the modeling of a GoogLeNet Convolutional Neural Network (CNN) on the GPC architecture and evaluates its performance and scalability. The models feature a new modular Memory Access Resources and Interfaces (MARI) library to improve communication between modules and assist during profiling. This work also introduces a graphical CAD software called Map Grid-based Layouts (MapGL) to facilitate the design process, automatically generate SystemC models, and generate performance reports. Experimental results evaluate and compare the generated models and show the achieved improvements in terms of memory usage and speed.

Relators: Guido Masera, Rainer Dömer
Academic year: 2022/23
Publication type: Electronic
Number of Pages: 74
Corso di laurea: Corso di laurea magistrale in Ingegneria Elettronica (Electronic Engineering)
Classe di laurea: New organization > Master science > LM-29 - ELECTRONIC ENGINEERING
Ente in cotutela: University of California Irvine (STATI UNITI D'AMERICA)
Aziende collaboratrici: University of California, Irvine
URI: http://webthesis.biblio.polito.it/id/eprint/24562
Modify record (reserved for operators) Modify record (reserved for operators)