On the Improved Performance of VLPL-GPU Plasma Simulation Code
摘要
Abstract
The GPU version of Q-VLPL3D code was improved by replacing the FFT Poisson solver by solver from MAGMA library. Also the Particle-In-Cell method implementation based on CUDA surfaces was rewritten using just ordinary GPU memory. This resulted in performance of 1.5 PFLOPS during particle push with Nvidia A100 and 4.5 TFLOPS during electromagnetic field evaluation.