cuGMEC: A high-performance code for gyrokinetic-MHD hybrid simulation on GPUs with CUDA C++
Description
Studying the interactions between energetic particles (EPs) and Alfvén eigenmodes (AEs) is essential for understanding alpha particle transport in burning plasma conditions within future fusion reactors. To numerically simulate EP-driven Alfvén instabilities, we have developed GMEC[1,2] (Gyrokinetic-Magnetohydrodynamics Energetic-particle Code) on the central processing unit (CPU) platform. However, long-time and large-scale simulations require higher grid resolution and more particles, demanding substantial computational resources. Therefore, applying more efficient numerical algorithms and optimization strategies on advanced computing architectures becomes critical. Computer scientists have designed the graphics processing unit (GPU) to implement single instruction multiple thread (SIMT) with fewer control units and higher throughput than CPUs. The GPU is optimized for data-parallel tasks and has emerged as a powerful platform for scientific computing. Based on NVIDIA’s Compute Unified Device Architecture (CUDA), this work presents cuGMEC, a high-performance gyrokinetic-MHD hybrid code developed from scratch for GPU platforms, incorporating newly added equations and physical terms. After testing on NVIDIA GPUs, the results show favorable scaling and acceleration, achieving up to 20 times speedup compared to the Intel Xeon Gold 6348 processor. Several benchmarks with theories and other codes have also been conducted to verify the gyrokinetic-MHD hybrid code.