GPU Accelerated MoM Matrix Inversion
The GPU Solver enables WIPL-D Pro to significantly decrease the simulation time, particularly for the dominant component – MoM Matrix inversion. This applies especially in the case of very complex and electrically large structures with dramatically large number of unknowns. In such cases, the suggested simulation approach is using GPUs to perform computing operations instead of CPU.
Significant acceleration is demonstrated even for the problems with 50,000 – 150,000 unknowns compared to a CPU workstation with 2 CPUs, each with 12 cores. Such problems are within reach of the CPU solution, especially if there is enough available RAM.
The app note shows the simulation times for an affordable workstation with 4 low-cost GPU cards and several hard discs supporting the GPU solver. All times are rather short and acceptable from the engineering point of view.
If the system matrix is symmetrical, as in the cases of pure metallic structures, the reduced matrix inversion can be applied with significantly shortened simulation time. The two demo problems are Luneburg lens and the Cassegrein reflector antenna. 440 lambda Cassegrein, normally out of reach for any full wave solver, is solved in only 6 hours for the 500,000 MoM unknowns problem.
For full version of the document, please check the following pdf.