New in 13.0 is the GPU Accelerator capability.

When using a supported graphics card (namely, NVIDIA Tesla cards) and having an ANSYS HPC Pack license, one can take advantage of the cores on the graphics card for the computation.  Unfortunately, this doesn’t work on any graphics card — for example, many graphics cards only use single-precision (ANSYS needs double-precision for accuracy), not enough memory is available on the graphics card, and the cores on the graphics card may not be fast enough to provide any improvement compared with the regular CPU cores.  However, if one has a supported graphics card, with the GPU Accelerator capability, calculations during the solution (namely, equation solution portion — not the element formulation or element results calculations) can be much faster.

For example, a harmonic response analysis of a vibroacoustic problem, which results in unsymmetric matrices, showed the following speed-up using GPU Accelerator capability (comparison is with same number of cores with vs. without GPU Acceleration — comparison is not using 1 core case as baseline):

  • 1 core – 4.2x faster with GPU Accelerator
  • 2 cores – 2.8x faster
  • 3 cores – 1.2x faster
  • 4 cores – 1.5x faster

In another case, a modal analysis with piezoelectric materials was performed, also using the sparse direct solver (this case results in symmetric matrices).  The speedup with GPU Acclerator was as follows:

  • 1 core – 3.3x faster
  • 2 cores – 2.0x faster
  • 3 cores – 1.6x faster
  • 4 cores – 1.4x faster

Results will vary, depending on the hardware configuration and problem solved, by one can see good speedup with the above cases.  Not only does one get speedup by using more cores (speedup not listed here), but GPU Accelerator further increases the performance.  (Actually, these cases were not run on a Tesla card but an NVIDIA Quadro 6000 card.)  Additional benchmarks can be found in this article from the ANSYS Advantage publication.

More details on GPU Accelerator can be found in the ANSYS 13.0 Help system, under the Mechanical APDL “Advanced Analysis Guide”, Chapter 16 “GPU Accelerator Capability”.  Note that matrices with Lagrange Multipliers (e.g., MPC184, contact with Lagrange Multipliers, Mixed u-P formulation) do not utilize GPU Accelerator, although unsymmetric matrices (e.g., vibroacoustics, frictional problems, nonassociative plasticity, etc.) are supported.