GPU accelerated EM modelling in frequency domain: Comparison of performance of various GPU cards

In this paper, we compared performances of different NVIDIA CUDA-enabled GPUs in accelerating MoM based software for EM modelling in frequency domain, WIPL-D. Three different GeForce GPUs were used: GTX 480, GTX 580 and GTX 680. The performance was measured on three EM problems: PEC sphere, Luneburg lens, and microstrip patch antenna array placed at the helicopter fuselage. The best performance in matrix fill was achieved on GTX 580, and the best performance in matrix inversion was achieved on GTX 680. Also, it was shown that the most critical parameter for achieving better performance is GPU memory bandwidth.