FPGA-based Custom Computing Architecture for Large-Scale Fluid Simulation with Building Cube Method

Abstract

We are designing a custom computing machine for large-scale flui simulation with the building-cube method (BCM). In BCM, parallel computation is performed with cubes, each of which is an orthogonal grid with a f xed resolution of cells. Although BCM is advantageous in balancing loads with cubes, it also has a problem of efficien y and scalability for comptuting with general-purpose supercomputers due to insufficien memory bandwidth and communication overhead of an interconnection network. In this paper, we present a custom computing architecture for FPGA-based scalable BCM computation with a dedicated network, called an accelerator domain network (ADN). We design a cube engine which allows bandwidth-efficien computation of cubes based on streamed stencil computation of the fractional-step method. Through prototype implementation, we evaluate the potential performance of the architecture. For ALTERA Stratix V 28nm FPGA, we estimate that a single FPGA has the peak performance of 107 GFlop/s in a single precision.