Machine learning accelerator design in HDL
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
With the development of computer science, the demand for target recognition using convolutional neural networks has increased significantly. However, traditional computing platforms such as CPUs and GPUs often suffer from limitations including slow processing speed, high power consumption, and limited flexibility in certain embedded or specialized scenarios. This project focuses on the implementation of an HDL-based neural network accelerator and employs a golden reference model for functional validation. The project implements the classic LeNet-5 architecture, consisting of two convolutional layers, two pooling layers, two fully connected layers, and one SoftMax layer, for a total of seven layers. To balance processing speed and hardware resource utilization, parallel computing and pipeline processing techniques are adopted in the convolutional layers. For system validation, the MNIST dataset was trained using PyTorch, achieving an accuracy of approximately 99%. The convolution and pooling layers were implemented in Verilog, while the fully connected and output layers were realized in Matlab for post-processing and verification of the hardware simulation results. Experimental results show that the proposed accelerator achieves an inference latency of about 20 μs for LeNet-5 on the MNIST dataset.
Fields of Study & Subjects
Embedded Systems Design Techniques
Advanced Neural Network Applications
Embedded Systems and FPGA Applications
Publication details
- Journal / source
- DR-NTU (Nanyang Technological University)
- Publication date
- 1 January 2026
- Language
- EN
- Document type
- Thesis