A High Performances and High Versatility Reconfigurable System for Fast Prototyping of Digital Neural Network Based on FPGA

Abstract

In this paper we present the design and the realization of a system suitable to work as a fast prototyping platform for implementing several digital neural networks by means of FPGA each one with a different number of processing units and connectivity architecture. Thus we developed a system with a processing kernel working on neural algorithms in addition with a control unit and a communication unit that allow the managing and the interfacing both toward application field and a host computer for data downloading. To get the same performances varying the neural network implemented we developed a communication protocol that handle all the working operations. Its generality and adaptability at various working conditions allow an easy reconfiguration of the features of the neural network and therefore a saving in the time spent in testing and prototipation. Then we will show how a neural network can be implemented using up to four FPGA depending on its complexity. A case study will be presented to show the performances of the overall system.