Fast and compact evolvable systolic arrays on dynamically reconfigurable FPGAs

Abstract

Evolvable hardware may be considered as the result of a design methodology that employs an evolutionary algorithm to find an optimal solution to a given problem in the form of a digital circuit. Evolutionary algorithms typically require testing thousands of candidate solutions, taking long time to complete. It would be desirable to reduce this time to a few seconds for applications that require a fast adaptation to a problem. Also, it is important to consider architectures that may operate at high clock speeds in order to reach very speed-demanding situations. This paper presents an implementation on an FPGA of an evolvable hardware image filter based on a systolic array architecture that uses dynamic partial reconfiguration in order to change between different candidate solutions. The neighbor to neighbor connections of the array offer improved performance versus other approaches, like Cartesian Genetic Programming derived circuits. Time savings due to faster evaluation compensate the slower reconfiguration time compared with virtual reconfiguration approaches, but, at any rate, reconfiguration time has been improved also by reducing the elements to reconfigure to just the LUT contents of the configurable blocks. The techniques presented in this paper lead to circuits that may operate at up to 500 MHz (in a Virtex-5), filtering 500 megapixels per second, the processing element size of the array is reduced to 2 CLBs, and over 80000 evaluations per second in a multiplearray structure in an FPGA permit to obtain good quality filters in around 3 seconds of evolution time.