Computational acceleration of ICF capsule simulations through GPU-enabled plasma-physics modules

This work will detail efforts to speed up computational simulations of ICF capsules on GPU-based computing environments such as the Sierra/Lassen machine (NVIDIA GPUs) and the upcoming El Capitan exascale supercomputer (AMD GPUs). Marbl is a high-energy-density-physics solver developed at LLNL for various applications including laser fusion and pulsed power, and it relies on curvilinear high-order finite-element methods for ALE hydrodynamics [1]. We will highlight the steps followed to achieve fast GPU performance of the plasma-physics modules in Marbl (3T physics, thermonuclear burn, electron thermal conduction, etc.). Metrics such as strong, weak, and throughput scalings will be compared for simulations on standard CPU-only architectures (Intel Broadwell) and GPU-enabled heterogenous machines (NVIDIA’s V100 and AMD’s MI250X). The two test cases used for this analysis are a pre-imploded hotspot and the Tipton capsule [2].

[1] R. W. Anderson et al., “High-order multi-material ALE hydrodynamics,” SIAM J. Sci. Comput. Vol. 40, No. 1, pp. B32-B58, 2018.

[2] R. Tipton, “An ICF test problem for single group radiation-hydrodynamics codes”, unpublished report.

This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344.