Hardware Implementation of an Artificial Neural Network with an Embedded Microprocessor in a Fpga

This article describes the implementation in hardware of an Artificial Neural Network with an embedded Microprocessor in a FPGA. The implementation of a Neural Network in hardware can be desired to benefit from its distributed processing capacity or to avoid using a personal computer attached to each implementation. The relevance of implementing it in a FPGA comes from its flexibility, low power consumption and higher performance. This implementation uses an embedded processor. Embedding the processor allows achieving the benefits from hardware and from software in a single platform. The implementation with a microprocessor can be quite easy while a regular hardware implementation can be hard to develop. The hardware implementation is based in a Feedforward Neural Network, with a hyperbolic tangent as activation function, with floating point notation of single precision. The device used was an FPGA Virtex II Pro XC2VP30, Xilinx with a MicroBlaze soft core processor. The microprocessor soft core subsystem occupied 1766 slices and it used 89KB of RAM for the application, data and results storage (the base value of on-chip memory is of 64KB). The Matlab was used to validate the implementation by comparing it with the results of the hardware solution while using data from a real system. The results show that the implementation does not introduce a noticeable loss of precision but is slower than the Matlab implementation running in a PC with a processor running at 2,8GHz.