Can a GPU be used to accelerate the FDTD?

Altair Forum User
Altair Forum User
Altair Employee
edited October 2020 in Community Q&A

With the FDTD solver a GPU can be used to accelerate the solution. Could someone provide more details on the conditions for using the GPU and also some tips?

Tagged:

Answers

  • JIF
    JIF
    Altair Employee
    edited June 2017

    The FDTD method is well suited for GPU acceleration. The GPU acceleration is supported for Windows and Linux platforms. Note that currently only NVIDIA GPU devices are supported by the FEKO kernel (since FEKO Suite 6.0). More details regarding the list of supported NVIDIA graphic cards can be found at http://www.nvidia.com/object/cuda_gpus.html. It is recommended that users install the latest NVidia drivers (http://www.nvidia.com).

    How many GPUs can be used and can they be used in parallel:

    FEKO presently supports the use of a single GPU in a sequential solution for the FDTD. Multiple GPUs and parallel solutions are in development.

    Memory requirements:

    When the GPU is used, the entire model must fit into the GPU memory. If the FDTD model requires more memory than available on the GPU, FEKO will deactivate GPU solving and solve the model with the CPU only. In general, the GPU pricing increases with the available memory of the GPU. A good compromise between price and performance is a GPU with 2 GByte of memory. This will solve a variety of small to medium sized FDTD problems. Electrically large models usually solve more efficiently with FEKO's frequency domain solvers, therefore users would not easily need more than 2 GByte of GPU memory.

    A recommendation is to have two GPUs - one entry level GPU used for the workstation's display, and a second GPU, possibly a more high end model, exclusively used for FEKO.

     

     

Categories