Solve Ax=B on an FPGA

• Main Author: Ling, Jun Han
• Other Authors: Ling Keck Voon
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; FPGA
• Online Access: https://hdl.handle.net/10356/176485
• Institution: Nanyang Technological University

This report describes the development and analysis of a hardware-accelerated Gaussian Elimination linear solver implemented on the PYNQ-Z1 Field Programmable Gate Array (FPGA) System on Chip (SOC). The primary intention was to explore the potential of FPGAs to accelerate computational tasks specifically mathematical calculations traditionally performed by software and leverage the parallel processing capabilities of the PYNQ Z1 platform. The project also intends to address and overcome the challenges and limitations of hardware acceleration specifically the memory transfer bottlenecks which is known source of throughput limitation for hardware acceleration. By presenting a comprehensive overview of the project’s design and implementation process, along with the encountered obstacles, with a detailed exploration of the problem, including the potential causes and workarounds. Despite the promising capabilities of FPGAs in enhancing computational speed, the project encountered a significant obstacle during the final stages of implementation, with the Direct Memory Access (DMA) integration. Through a detailed examination of the project’s design, this report discusses the encountered memory transfer bottleneck, emphasizing on the DMA-related challenge that posed a critical impediment. Additionally, it explores the theoretical solutions and workarounds for overcoming such limitations in FPGA-based projects. The project concluded with a successful design and implementation of a 128 by 128 floating point Gaussian Elimination solver implemented onto FPGA hardware, the PYNQ Z1, that processed the input data correctly and by measurements of the processing function’s run time, outperformed a software solution running on the same processing system.