Nvidia CUDA parallel processing of large FDTD meshes in a desktop computer
Robert Calatayud, Enrique Navarro-Modesto, Enrique A. Navarro, Nagula Sangary
Abstract
Robert Calatayud, Enrique Navarro-Modesto, Enrique A. Navarro, Nagula Sangary
Abstract
The Finite Difference in Time Domain numerical (FDTD) method is a well know and mature technique in computational electrodynamics. Usually FDTD is used in the analysis of electromagnetic structures, and antennas. However still there is a high computational burden, which is a limitation for use in combination with optimization algorithms. The parallelization of FDTD to calculate in GPU is possible using Matlab and CUDA tools. For instance, the simulation of a planar array, with a three dimensional FDTD mesh 790x276x588, for 6200 time steps, takes one day -elapsed time- using the CPU of an Intel Core i3 at 2.4GHz in a personal computer, 8Gb RAM. This time is reduced 120 times when the calculation is parallelized and carried out in a Graphics Processing Unit (GPU) NVIDIA GeForce GTX 1080 Ti 11264 MB GDDR 5X. The elapsed time is reduced substantially, but also the simplicity of calculation and usefulness of a Matlab implementation. The elapsed time reduction is so substantial that the FDTD-Matlab-CUDA can be combined with optimization algorithms.
OpenAlex reports 2 citations for this work. Citation counts describe recorded attention and do not establish research quality.
A contribution statement is not available in the OpenAlex record.
Method details are not available in the OpenAlex metadata.
Findings are not separately available in the OpenAlex metadata.
Limitations are not available in the OpenAlex metadata.
Application details are not available in the OpenAlex metadata.
The Finite Difference in Time Domain numerical (FDTD) method is a well know and mature technique in computational electrodynamics. Usually FDTD is used in the analysis of electromagnetic structures, and antennas. However still there is a high computational burden, which is a limitation for use in combination with optimization algorithms. The parallelization of FDTD to calculate in GPU is possible using Matlab and CUDA tools. For instance, the simulation of a planar array, with a three dimensional FDTD mesh 790x276x588, for 6200 time steps, takes one day -elapsed time- using the CPU of an Intel Core i3 at 2.4GHz in a personal computer, 8Gb RAM. This time is reduced 120 times when the calculation is parallelized and carried out in a Graphics Processing Unit (GPU) NVIDIA GeForce GTX 1080 Ti 11264 MB GDDR 5X. The elapsed time is reduced substantially, but also the simplicity of calculation and usefulness of a Matlab implementation. The elapsed time reduction is so substantial that the FDTD-Matlab-CUDA can be combined with optimization algorithms.
Key concepts: Finite-difference time-domain method, CUDA, Computer science, Computational science, MATLAB, Graphics processing unit, Parallel computing, Polygon mesh