Advanced Architectures for Astrophysical Supercomputing

TL;DR

This paper analyzes astronomy algorithms to identify optimal problem types and techniques for leveraging current and future GPU architectures, aiming to enhance astrophysical computations and scientific discovery.

Contribution

It provides a generalized approach to understanding and optimizing astronomy algorithms for advanced GPU architectures, facilitating better utilization of high-performance computing resources.

Findings

GPU speed-ups of up to 100x for general-purpose computation

Identification of problem types best suited for GPU acceleration

Guidelines for adapting algorithms to future architectures

Abstract

Astronomers have come to rely on the increasing performance of computers to reduce, analyze, simulate and visualize their data. In this environment, faster computation can mean more science outcomes or the opening up of new parameter spaces for investigation. If we are to avoid major issues when implementing codes on advanced architectures, it is important that we have a solid understanding of our algorithms. A recent addition to the high-performance computing scene that highlights this point is the graphics processing unit (GPU). The hardware originally designed for speeding-up graphics rendering in video games is now achieving speed-ups of $O(100\times)$ in general-purpose computation -- performance that cannot be ignored. We are using a generalized approach, based on the analysis of astronomy algorithms, to identify the optimal problem-types and techniques for taking advantage of both current GPU hardware and future developments in computing architectures.