Implementation and evaluation of a particle mover using mixed precision arithmetic

Abstract: Computer simulations are broadly used nowadays in order to obtain information that would be impossible to gain otherwise. These computational workloads have grown in size to the point where even a small improvement in its implementation can lead to a substantial speed-up. For this reason, researchers have studied the impact linked to the usage of less precise numbers, since it would accelerate calculations. Likewise, this study aims to determine the feasibility of using mixed-precision arithmetic in iPIC3D, a 3D implicit Particle-in-cell (PIC) implementation used in plasma simulations. Specifically, the use of mixed-precision numbers will be limited to the particle mover, the section that solves the equations of motion for each particle of the plasma, resulting in the most time-consuming part of the code. The results show a maximum divergence or error of about 2% between the original implementation and the new one when comparing the output values. All of this, performing a relatively short simulation of 2250 cycles, therefore, with longer tasks we the error could increase. Thus, we come to the conclusion that in most cases, the loss in precision is too high to justify the use of this new implementation.