Performance, accuracy, power consumption and resource utilization analysis for hardware / software realized Artificial Neural Networks

Artificial Neural Networks (ANN) are used to perform tasks like classification, pattern recognition and function approximations in many cases to which traditional approaches are not well suited. Hardware implementations have been presented, mainly in academical works, in order to take advantage of the inherent parallelism in ANNs. In the field of embedded systems it is desirable to have faster and less power demanding designs. This work analyzes implementations of ANNs in FPGAs both in Hardware Description Language (HDL) and in software code running on different configurations of the Xilinx MicroBlaze microprocessor. Three versions of an ANN design were implemented in HDL and a software version was executed in four different configurations of the Xilinx MicroBlaze microprocessor. Results for power consumption, FPGA occupation, speed and accuracy of the outputs are presented for practical experiments performed in two FPGAs from different families of Xilinx devices: a Spartan 3E and a Virtex 5.