Accelerating Numerical Simulations of Supernovae with GPUs

To understand the mechanism of supernova explosions, large-scale numerical simulations are essential because of their complex dynamics described by a coupled equations of neutrino radiation transport and hydrodynamics of dense matter. In this work, we employ GPUs to accelerate such simulations. By adopting the implicit scheme for the evolution equation, an iterative linear equation solver for the coefficient matrix is the most time consuming part, which has been shown to be efficiently offloaded to GPUs. There are still several secondary bottlenecks which cost substantial time in the simulations, such as computation of the collision term of the Boltzmann equation of neutrinos, and parameter tuning of the matrices in the iterative solver. This paper focuses on these parts and offloads them to GPUs by employing CUDA in the case of spherically symmetric system. As a result, the time evolution is sufficiently accelerated for desirable model sizes toward systematic survey of stellar models with better grid resolution than that adopted so far.