Performance of ML-Based Bandwidth Compression on FPGAs

Aleko Lilius ()

Abstract

This thesis investigates the integration of machine learning (ML)-based compression on Field-Programmable Gate Arrays (FPGAs) to enhance bandwidth compression of data, a crucial aspect in scientific research where large amounts of data are produced in real-time. The compression tool Baler, utilizing autoencoders for ML-based compression, is designed to handle scientific data efficiently. By combining the adaptability of ML models with the computational efficiency of FPGAs, this thesis aims to evaluate the performance of Baler's bandwidth compression. The thesis work reveals that smaller models can effectively fit onto the FPGA, resulting in a throughput increase of 16.9 times compared to a CPU in a desktop computer. This significant improvement demonstrates the potential of FPGA-accelerated ML solutions. Key factors influencing optimal FPGA performance, including model size, precision levels, and clock period, were identified. This thesis lays a foundation for further developing hardware implementation of the Baler algorithm, suggesting that the convergence of ML and FPGA technology holds significant potential for enabling more efficient hardware-accelerated ML solutions.