On the Hardware-Relevant Simulation of Regular Two-Dimensional CNN Processing Grids

Massively parallel processing architectures mimicking biological structures and their underlying calculation principles have been put into practice by the members of the cellular neural network (CNN) community. But until now flexible, scalable and industrially qualified toolkits are not available to support the simulation and development of these architectures within one single environment. In this paper we report on a simulation-framework, which is conceptualized and adjusted to deal with the specific simulation requirements of purely digital CNN processing devices. In particular, the framework is able to (1) handle complete CNN architectures of industrial relevant size, (2) to represent double precision float-point numbers as well as hardware relevant fixed-point numbers and (3) offer simulation run-times a magnitude faster than standard digital hardware simulations. We conclude this paper by presenting selected simulation results manifesting the proposed capabilities of the simulation-framework.