This thesis presents the development and implementation of neural network accelerators on FPGA boards, with a particular focus on the open-source hls4ml tool. The work covers fundamental concepts of neural networks, FPGA architecture, and optimization techniques such as quantization, weight pruning, and reuse factor, which reduce resource utilization and energy consumption. As a practical example, a convolutional neural network for classifying images from the MNIST dataset was implemented on the PYNQ-Z1 development board. Results demonstrate that the FPGA implementation achieves high throughput (938 images/s) and low latency (0.508 ms) while maintaining high classification accuracy (96.6\%). The thesis illustrates the complete workflow, from training the model in Python, converting it into an HLS project with hls4ml, to implementing and executing it on hardware.
|