Modern Parallel Architectures to Speed Up Numerical Simulation

Abstract

Applications of graphics processing units (GPU) and field programmable gate array (FPGA) for computer codes acceleration are discussed. Most of the high positions in the top-100 list of supercomputers (clusters) are taken by a hybrid type hardware. First, the authors provide an idea about GPU and FPGA architectures. The use of FPGA has two main obstacles, involving the necessity for manual coding of algorithms up to the register transfer level (RTL). So, modern high level synthesis (HLS) technology to use FPGA is briefly introduced. Then several examples of speeding up algorithms mostly from the Earth Sciences are given. The considered examples of GPU use are: decomposition of seismic records by wave packages (performance gain of 350 times is achieved); the convolution problems with Green’s function (the computation time at single GPU is 162 times faster than the original code version); and tsunami wave propagation (simulation of tsunami wave propagation was accelerated 100 times compared to one CPU). In some cases FPGA shows even better results compared to GPU, in particular for tsunami modelling (five times faster than compared even to GPU Tesla K40), HD video stream processing. As for FPGA-based data processing, the following examples are here considered: searching for small objects on a series of images; searching object on the image; and motif search in DNA sequence. In all cases comparison with one CPU is given.