GPU Acceleration of the PWTD Algorithm for application in High-Frequency Communication and Fotonics
R.J. Gravendeel (TU Delft - Electrical Engineering, Mathematics and Computer Science)
K. Cools – Mentor (TU Delft - Numerical Analysis)
K. Vuik – Graduation committee member (TU Delft - Numerical Analysis)
HX Lin – Graduation committee member (TU Delft - Mathematical Physics)
R. van Driel – Graduation committee member (Capgemini)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
When creating electronic devices, it is essential to model what happens when an electromagnetic field hits the device and it scatters. Conventionally, this can be modelled using the Marching-on-in-Time algorithm. This can become computationally expensive for complex systems. To speed up the algorithm, the Plane-Wave Time-Domain algorithm is combined with the MOT algorithm. To accelerate the process even more, part of the algorithm is implemented using a Graphics Processing Unit, or GPU.
To test if using GPUs for this type of problem is actually beneficial, three experiments are set up. The first one tests the basic operations of addition and multiplication on matrices and vectors of various sizes, to determine if and when the computation time of the GPU is lower than that of a CPU. The second experiment tests the use of Fast Fourier Transform planner functionality and compares the CPU computation time with that of the GPU for the FFT of matrices of various sizes. The third experiment compares an example of the PWTD algorithm on the CPU and the GPU. These experiments are performed on three different devices.
The results from experiment 1 and 2 show that, after a certain point, the GPU is almost always faster, no matter the operation. Experiment 3 shows that the current GPU implementation is currently not as fast as the regular PWTD algorithm, though one of the devices is only 0.003\% slower.
In conclusion, theoretically a decrease in computation time is expected. From experiment 3 it follows that it is not the case yet, though with more optimisation the GPU implementation would almost certainly become faster.