Practical acceleration of direct whole-core calculation employing graphics processing units

Abstract

The methods and performance of GPU-accelerated Direct Whole-Core Calculation (DWCC) are presented. GPU computing techniques were introduced to significantly reduce the execution time of the production-grade DWCC code nTRACER that employs the planar Method of Characteristics (MOC) and augmented axial MOC within the framework of the Coarse Mesh Finite Difference (CMFD) acceleration. The data structures and algorithms are largely modified from the legacy CPU solver in a way to be suitable for the computational characteristics of GPUs. Mixed precision technique is extensively utilized in order to exploit the superior single precision computing power of consumer-grade GPUs and to make use of limited GPU memories efficiently. Concurrent execution of CPU and GPU is then utilized to take advantage of heterogeneous computing architectures. This includes the source update by CPUs for lower energy groups during the ray tracing by GPUs for upper energy groups. The performance examination indicates that the planar MOC calculation time can be remarkably reduced by GPU acceleration. As the result of efficient GPU acceleration, a three-dimensional (3D) DWCC calculation for a typical light water reactor could be carried out in a few minutes on an industrially affordable cluster mounted with consumer-grade GPUs.