Performance Benchmark of Cahn–Hilliard Equation Solver with Implementation of Semi-implicit Fourier Spectral Method

The performance scaling issue of phase-field simulation is one that must be overcome to perform realistic large-scale three-dimensional prediction. The CUDA (Compute Unified Device Architecture) parallel acceleration method developed over a decade ago showed very good performance in terms of calculation speed, but was limited by the small size of memory on the GPU. Recently, Apple Inc. has announced a GPU–CPU hybrid architecture, Apple silicon (M1 or later), and we examine the advantages of this architecture for performing realistic large-scale phase-field simulations and compare it to existing CUDA architecture. When solving the Cahn–Hilliard equation using the FFT (Fast Fourier Transform) with CUDA architecture developed by Nvidia and Apple silicon architecture developed by Apple Inc., we compared performance across hardware, as well as other considerations such as form factor and heat dissipation of the workstation.