Introduction of GPU-Enabled CMFD Acceleration for Performance Enhancement of Neutron Transport Code STREAM3D-GPU

Abstract

A GPU-enabled STREAM3D-GPU code has been developed at UNIST to enable large-scale reactor analysis for a single burnup step in the 3D MOC/DD transport framework within a few minutes. In addition to previously offloaded MOC sweeping in this work, a GPU-enabled CMFD acceleration was implemented, with the latter being our primary focus in the current study. The code has been verified using a 3D OPR-1000 octant-core depletion problem with thermal-hydraulic feedback over 30 burnup steps. Results demonstrate the GPU implementation's effectiveness, reducing the runtime to 32 minutes per burnup step while keeping the multiplication factor within 10 pcm. Root mean square percentage error (RMSPE) for pin powers stayed below 0.04% for all depletion points, and RMSPE for pin burnups was found within 0.02% for all measured core burnups. This result has been achieved using a single commercially available GPU node that contains eight consumer-grade GPUs and a 64-core CPU.