Academia Sinica Logo

Abstract

Until the early 2000s, uniprocessor performance climbed along with Moore's law, doubling approximately every two years. To obtain further speedups with computer programs, one had to only purchase a new machine. However, microprocessor manufacturers hit a fundamental wall, which meant that clock frequencies could not be increased further with mass-produced processor coolers. After this turning point, sequential programs had to be parallelized to obtain further speedups. The free lunch was over. Today, graphics processing units are the flagships of massively parallel accelerators, which have been used successfully to accelerate machine learning, computational chemistry, and physics by an order of magnitude compared to general-purpose purpose processors traditionally used in high-performance computing. In this talk, we will introduce the concepts of implementing physical simulation software on graphics processing units, discuss the outlook of the high-performance computing landscape, and give an overview of the state-of-the-art methods for accelerating astrophysical simulations.