Your browser does not support JavaScript!

Home    Collections    Type of Work    Post-graduate theses  

Post-graduate theses

Current Record: 10 of 4831

Back to Results Previous page
Next page
Add to Basket
[Add to Basket]
Identifier 000429201
Title Design-space exploration of FPGA architectures for efficient HPC acceleration
Alternative Title Εξερεύνηση αρχιτεκτονικών προγραμματιζόμενης λογικής για αποδοτική επιτάχυνση εφαρμογών υψηλών επιδόσεων
Author Γιαουρτάς, Μιχαήλ Π.
Thesis advisor Κατεβαίνης, Μανώλης
Reviewer Παπαευσταθίου, Βασίλης
Πρατικάκης, Πολύβιος
Δόλλας, Απόστολος
Abstract Field Programmable Gate Arrays (FPGAs) have an ever-expanding impact to more and more applications, ranging from Deep Neural Networks to High-Performance Computing (HPC) and other uses such as customization of Instruction Set Extensions and computation offloading in systems with tightly coupled embedded FPGAs (eFPGAs). As applications diverge in complexity, performance, memory needs and area limitations, there is a need for a wider variety of FPGA architectures. However, developing and implementing new FPGA architectures remains challenging and requires a lot of time, due to their high content in custom layout designs and the need for design software and flows tailored for each specific architecture, leading to the production of more generic products. Many academic works are focusing on the automated FPGA design generation process, in an attempt to promote customizability and reduce time-to-market. In other approaches, researchers target only the exploration process, in which they seek for the optimal architecture for a specific case scenario, using area and delay estimation models. In this thesis we choose to combine the two approaches. We develop an extension for the popular open-source tool Verilog-to-Routing (VTR) in order to export in Verilog the representation of user-specified FPGA architectures, develop support for custom user hard-blocks (RAMs, DSPs, FP Units), and generate Bitstreams for given benchmarks. Our objective is to create synthesizable and technology independent RTL design code, able to be synthesized with any standard cell library. We discover the real design properties of an FPGA architecture using our proposed ASIC flow and retrieve real area and delay measurements and eventually proceed with the exploration of optimal FPGA architectures for given sets of benchmarks. Using our VTR extension, we perform FPGA design-space exploration for a set of HPC oriented benchmarks that are derived using Xilinx's High Level Synthesis (HLS). Our exploration starts by identifying pareto-optimal FPGA architectures starting with the size of Lookup Tables (LUTs) and the number of LUTs per Configurable Logic Block (CLB) and then explore the size of routing channels and wire segments' configurations. We also compare the optimal FPGA architectures derived when using the HPC benchmarks with the respective architectures derived when we use the generic MCNC benchmarks. Finally, we create TCL scripts for synthesis and back-end implementation (place and route) which can adjust to any architectural characteristic and size and automate the ASIC flow for new FPGA chips.
Language English
Subject ASIC
Προγραμματιζόμενη λογική
Issue date 2020-03-27
Collection   Faculty/Department--Faculty of Sciences and Engineering--Department of Computer Science--Post-graduate theses
  Type of Work--Post-graduate theses
Permanent Link Bookmark and Share
Views 864

Digital Documents
No preview available

No permission to view document.
It won't be available until: 2020-09-27