Core Concepts
Efficient GPU implementation of spectral-element methods for fast solving of 3D Poisson equations.
Abstract
The content discusses the development of a simple GPU implementation of spectral-element methods for solving 3D Poisson equations on rectangular domains. It explores the tensor product structure of Laplacian on Cartesian meshes and presents a MATLAB implementation for solving 3D Poisson equations using a spectral-element method. The article highlights the successful GPU acceleration of numerically solving PDEs and its applications to linear and nonlinear equations. It also emphasizes the importance of fast solvers for Poisson equations in various scientific and engineering fields.
The article is structured as follows:
Introduction to the tensor product structure of the discrete Laplacian.
Implementation details for 3D problems and high-order elements.
Inversion methods using eigenvalue decomposition.
Robust computation of the generalized eigenvalue problem.
Numerical tests and comparisons with FFT on GPU.
Application to solving a Cahn-Hilliard equation.
Stats
It costs less than one second on a Nvidia A100 for solving a Poisson equation with one billion degree of freedoms.
Quotes
"It is well known since 1960s that by exploring the tensor product structure of the discrete Laplacian on Cartesian meshes, one can develop a simple direct Poisson solver with an O(N^d+1/d) complexity in d-dimension."
"We present in this paper a simple but extremely fast MATLAB implementation on a modern GPU, which can be easily reproduced, for solving 3D Poisson type equations using a spectral-element method."