Methods and performance of a GPU-based pinwise two-step nodal code VANGARD

Abstract

The methods and performance of a GPU-based pinwise two-step nodal core calculation code VANGARD are presented which is targeted for use in the commercial nuclear design analyses. In order to realize high-speed core cycle depletion with a single consumer-grade GPU mounted on ordinary personal computers, various elaborated computational methods are introduced as follows. The computationally intense pinwise nodal calculation is optimized by adaptive employment of low-order expansion. The two-level surface current correction method is used for updating pinwise incoming partial currents within the assembly-level coarse mesh finite difference framework. A cross section data compression technique is introduced to fit the enormous pinwise group constant data into the limited GPU memory. A massively parallelized depletion scheme based on the Chebyshev Rational Approximation Method is implemented. The neighbor-informed burnup correction method is developed to ensure the accuracy of the pinwise depletion for gadolinia-bearing pins. Performance examinations are performed for the realistic APR1400 and AP1000 core problems to demonstrate that three-dimensional pinwise cycle depletion can be completed within 3 minutes with high accuracy and resolution.