Lessons from accelerating an RBF-FD phase-field model of dendritic growth on GPUs

Abstract

Abstract Phase-field modeling of dendritic growth presents the state of the art in the field of solidification modeling and are usually implemented using finite difference models combined with explicit time marching and accelerated by using GPUs. They are a prime candidate for such acceleration, since they require many arithmetic operations on relatively low ammount of data. We present an attempt at porting an existing RBF-FD code optimized for CPU execution to use GPU acceleration while keeping the resulting implementation portable between architectures. We discuss the acceleration achieved, scaling and implementation issues and critically discuss current landscape of GPGPU offerings.