Optimization of a VHDL Open-Source GPU Core: Analysis and Architecture Enhancement for Deep Learning Applications

This thesis describes the architecture and the enhancement process of an open-source soft-GPU for FPGAs (FGPU) for deep learning applications. Initially, an extensive study has been conducted to investigate the state-of-the-art of the available embedded GPU solutions. Thereafter, the FGPU has been chosen as a promising architecture to enhance. Due to lack of an accurate documentation, a relatively big effort has been made in the reverse engineering of the architecture. Eventually, the FGPU has been enhanced for more compliance with the OpenCL standard, which in the original version was not fully supported. This increase of compatibility has led to a reduction of the programming effort (and consequently of the development time) while simultaneously providing new functionalities and better performances