Simulation of three-dimensional supersonic flow on IBM Cell processor based emulated digital Cellular Neural Networks

Abstract

The IBM Cell Broadband Engine (Cell BE or informally Cell) heterogeneous multi-core architecture provides high computing performance however its efficient programming is a challenging task. The IBM Cell architecture can be very efficiently used to emulate Cellular Neural Network (CNN) dynamics. Using a similar approach during the simulation of two-dimensional supersonic flows significant speedup was achieved compared to conventional multi-core architectures. To solve real life flow problems three-dimensional mesh is required for the simulation. In the paper the implementation possibilities of a three-dimensional supersonic flow simulator on Body Fitted Mesh geometry on the Cell is described. Significantly more local memory is required to generate the three-dimensional neighborhood of the grid points compared to the two-dimensional case. In the paper different partitioning methods will be described and compared in terms of computing performance, local memory and bandwidth requirements.