講師：　Dr. Harald Koestler (University Erlangen-Nuremberg)

Abstract

Many real world applications involving numerical algorithms require high performance computing, either because of time constraints like for real-time simulations or because of a large amount of data to be processed.

GPUs offer high computational performance at low cost and are therefore an interesting architecture especially for data parallel applications. Recently, more and more GPU HPC clusters arise and thus there is a need for adapting numerical codes to Multi-GPU environments. In this talk we discuss software development for these clusters, compare CUDA and OpenCL, and show performance results on different GPU platforms and clusters.

As applications we consider two numerical algorithms used e.g. in computational fluid dynamics (CFD) or imaging. First we discuss 2D multigrid algorithms for image processing. Multigrid is among the most efficient numerical solvers for a variety of large, sparse (linear) systems arising from discretized partial differential equations (PDE). The goal here is to achieve maximum single GPU performance to enable real-time computations.

Second we address the Lattice Boltzmann Method (LBM) that follows in contrast to the Navier-Stokes equations a microscopic CFD approach based on cellular automata and kinetic theory. Here, due to the huge amount of data in the simulations multi-GPU support becomes mandatory and we will show scaling results for LBM on GPU clusters.