HyperFPGA: SoC-FPGA Cluster Architecture for Supercomputing and Scientific applications

Since their inception, supercomputers have addressed problems that far exceed those of a single computing device. Modern supercomputers are made up of tens of thousands of CPUs and GPUs in racks that are interconnected via elaborate and most of the time ad hoc networks. These large facilities provide scientists with unprecedented and ever-growing computing power capable of tackling more complex and larger problems. In recent years, the most powerful supercomputers have already reached megawatt power consumption levels, an important issue that challenges sustainability and shows the impossibility of maintaining this trend. With more pressure on energy efficiency, an alternative to traditional architectures is needed. Reconfigurable hardware, such as FPGAs, has repeatedly been shown to offer substantial advantages over the traditional supercomputing approach with respect to performance and power consumption. In fact, several works that advanced the field of heterogeneous supercomputing using FPGAs are described in this thesis \cite{survey-2002}. Each cluster and its architectural characteristics can be studied from three interconnected domains: network, hardware, and software tools, resulting in intertwined challenges that designers must take into account. The classification and study of the architectures illustrate the trade-offs of the solutions and help identify open problems and research lines, which in turn served as inspiration and background for the HyperFPGA. In this thesis, the HyperFPGA cluster is presented as a way to build scalable SoC-FPGA platforms to explore new architectures for improved performance and energy efficiency in high-performance computing, focusing on flexibility and openness. The HyperFPGA is a modular platform based on a SoM that includes power monitoring tools with high-speed general-purpose interconnects to offer a great level of flexibility and introspection. By exploiting the reconfigurability and programmability offered by the HyperFPGA infrastructure, which combines FPGAs and CPUs, with high-speed general-purpose connectors, novel computing paradigms can be implemented. A custom Linux OS and drivers, along with a custom script for hardware definition, provide a uniform interface from application to platform for a programmable framework that integrates existing tools. The development environment is demonstrated using the N-Queens problem, which is a classic benchmark for evaluating the performance of parallel computing systems. Overall, the results of the HyperFPGA using the N-Queens problem highlight the platform's ability to handle computationally intensive tasks and demonstrate its suitability for its use in supercomputing experiments.