مرکز منطقه ای اطلاع رساني علوم و فناوري - A CONTROLLER DESIGN TOOL DEVELOPMENT FOR AUTOMATICALLY MAPPING ARTIFICIAL NEURAL NETWORKS ONTO FPGAS

Abstract :

Artificial Neural Networks (ANNs) are the mathematical models which are based on neurons (brain cells) and the network of the neurons. Since ANNs require parallel distributed calculations, they usually are implemented on hardware when software implementations do not provide sufficient performance. Field Programmable Gate Array (FPGAs) chips are the best implementation option for ANNs due to their parallel processing and reconfiguration ability. On the other hand, implementing ANNs on FPGAs is a time consuming process and requires expert personal. Design and implementations made by humans are error prone and errors make the designers go back and reconsider their design from the beginning. In this study, ANNCONT has been developed as a part of a beforehand developed FPGA based automatic ANN design system. In this system, ANNCONT is responsible for designing a controller for a given ANN data-path and forming an ANN system by integrating this controller with the given data-path. ANNCONT has been tested with several test cases. Our observations show that it is able to design controllers and generate VHDL code for both the controller and the ANN system in less than a second without any errors in the code. Using ANNCONT, design and implementation processes can be shortened in terms of time, and expert requirement is minimized. Moreover, since ANNCONT produces error free code, debugging stage is eliminated.