A flexible multilayer perceptron co-processor for FPGAs

Abstract

Implementing a custom Artificial Neural Network (ANN) in hardware lacks the scalability and the flexibility of changing from one topology to another at run time. This paper presents a Multilayer Perceptron Co-processor (MLPCP) targeting FPGAs that is configurable during design time and programmable during run time. The MLPCP can be reprogrammed at run time to rapidly change network topologies and use different activation functions. This allows application developers to change parameters of a given network without the need to resynthesize. This also allows the MLPCP to be used for different applications during run time. Run time results show the MLPCP can deliver performance levels close to those of a custom ANN, and can execute network topologies that cannot fit into FPGAs with limited resources. Performance comparisons against software versions show up to 70x speedup compared to a MicroBlaze running at 100 MHz, and 4x compared to a Zynq running at 667 MHz.