A Stacking Scheme to Improve the Efficiency of Finite-Difference Time-Domain Solutions on Graphics Processing Units

摘要

Advances in computer hardware technologies accompanied by easy-to-use parallel programming software platforms have led to the wide spread use of parallel processing architectures, such as multi-core central processor units (CPUs) and graphic processing units (GPUs), in technical and scientific computing. Among electromagnetic numerical analysis methods, the finite-difference time-domain (FDTD) method is very well suited for parallel programming, and several implementations of FDTD have been developed and reported to solve electromagnetics problems orders of magnitude faster. Examination of performances of these implementations reveals that, in general, it is more efficient to solve larger FDTD domains than smaller domains. In this paper it is demonstrated that one can exploit the higher efficiency inherent to the solution of larger problem sizes to solve parameter sweep and optimization problems faster: instead of solving multiple smaller FDTD domains separately, these domains can be combined or stacked to form a larger problem and the large problem can be solved more efficiently. It has been shown that up to 40% faster solution can be achieved on GPUs with this method.