Hardware accelerators for neural networks: a review of the current state and prospects for development
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
This study is aimed to analyze the key aspects of hardware acceleration of neural networks and assess the effectiveness of the modern accelerator frameworks. FPGAs are chosen due to their high reconfigurability and parallelism in comparison to Von Neumann-inspired architectures. The research focuses on hardware-software co-design with the use of programmable logic devices. First, software optimizations are described, including pruning, distillation, quantization, and spiking neural networks as an efficient approach for processing sparse datasets. Common hardware design patterns are discussed, separating optimizations appropriate for dense and sparse inputs, including dataflow patterns and activation passing mechanisms. The research examines the existing FPGA-based accelerators for neural networks comparing their accuracy, hardware complexity, energy efficiency, and ability to perform on-chip training. Transfer learning is identified as a promising research area. Other types of programmable logic devices were analyzed, CPLD is identified as a potential energy-efficient target platform for deploying binarized neural networks. The practical value of the research lies in the potential application of the described principles for deploying neural networks on edge devices
Fields of Study & Subjects
Embedded Systems Design Techniques
Advanced Neural Network Applications
VLSI and FPGA Design Techniques
Publication details
- Journal / source
- COMPUTER-INTEGRATED TECHNOLOGIES EDUCATION SCIENCE PRODUCTION
- Issue
- 62
- Pages
- 55–66
- Publication date
- 28 March 2026
- Language
- EN