STREAM 3D GPU

Download Intro_STREAM 3D GPU (to be prepared)

The Recommended Publication for Citing

• 1. Siarhei Dzianisau, Deokjung Lee*, “GPU Acceleration of 3D MOC Solver in STREAM3D Using OpenACC,” KNS Spring Meeting, Jeju, South Korea, May 9-10 (2024).

  2. Siarhei Dzianisau, Muhammad Farid Khandaq, Tuan Quoc Tran, Deokjung Lee*, “On the GPU Acceleration of a 3D MOC Code STREAM Using OpenACC,” M&C 2023, Niagara Falls, Canada, August 13-17 (2023).

Introduction

STREAM3D-GPU is a GPU-enabled version of our in-house neutron transport code STREAM3D. A separate set of methods based on the original method of characteristics (MOC) solver of STREAM was developed and incorporated, which allowed STREAM to minimize GPU onboard memory occupancy. The current version of STREAM3D-GPU allows modeling a large PWR geometry such as OPR-1000 in less than 1 hour using a single GPU node. Unlike abundant in literature 2D/1D MOC codes, STREAM3D-GPU employs 72 neutron energy groups, fine radial mesh, and dense axial plane mesh (200 axial planes) for achieving the most accurate solution via our in-house 3D MOC/Diamond-difference (DD) method. The scalability of the code in terms of hardware was achieved because of the novel GPU memory management method, allowing the code to be used with any number of GPU cards available to the user regardless of the geometry settings. A set of techniques employed in STREAM3D-GPU ensures virtually linear scalability of the GPU code performance when adding or removing any number of GPU cards.

Features

• Support of all computation modes available in the CPU version of STREAM3D.

• High-fidelity calculation with thermal-hydraulic feedback and fine-mesh geometry.

• Linear scalability of GPU-enabled code.

• Efficient GPU memory management (3-4 GB per GPU card regardless of the number of GPU cards).

• The code is portable due to cross-platform OpenMP and OpenACC acceleration frameworks.

• Domain decomposition for single and multimode execution is available via MPI.

Developers

• Prof. Deokjung Lee: Supervision, funding

• Siarhei Dzianisau: Main developer

• Saisundar Mohanty: Developing transient solver

• Fathurrahman Setiawan: Developing CMFD acceleration

STREAM 3D CPU vs. GPU