Design of FPGA Accelerators based on CGRA Architectures

• Grant number: 24/08364-6
• Support Opportunities: Scholarships in Brazil - Post-Doctoral
• Start date: July 01, 2024
• End date: June 30, 2026
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computer Systems
• Principal Investigator: Guido Costa Souza de Araújo
• Grantee: Lucas Bragança da Silva
• Host Institution: Instituto de Computação (IC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• Associated research grant: 19/26702-8 - Trends on high performance computing, from resource management to new computer architectures, AP.TEM

Abstract

This project aims to accelerate complex scientific problems through FPGAs, using Coarse Grained Reconfigurable Arrays (CGRAs) architectures. CGRAs are architectures that can be specialized to run dedicated applications with a high level of parallelism. One of the limitations of CGRAs is that these architectures are, in most cases, synthesized on a single FPGA and cannot effectively deal with realistic scientific problems, which are orders of magnitude more complex than single-FPGA CGRAs can deal with. A-Machine is a Multi-FPGA cluster that was built within the scope of the FAPESP Thematic project "Trends in High-Performance Computing, from Resource Management to New Computer Architectures." It is located at the UNICAMP Institute of Computing, which is part of this thematic project. The A-Machine architecture has four nodes with two Alveo U55C FPGAs in each node, interconnected by fiber optic links. The A-Machine architecture allows kernels running on one of the FPGAs to communicate over optical links quickly and transparently with other kernels running on other FPGAs on other nodes. This is coordinated by the OpenMP Cluster (OMPC) runtime FPGA plugin, which was developed within the scope of the theme. This project will use the A-Machine and OMPC infrastructure to distribute many CGRAs among A-Machine FPGAs. This will expand the ability of CGRAs to deal with complex, realistic simulation problems, as it will considerably increase the spatial and temporal parallelism of their execution.