Learning Forum

[GeorgM]

Hi there,

the official HFSS documentation recommends the Intel Xeon Phi 7120A (MIC, not GPU) for the acceleration of the direct solver with the multi-frontal approach. However, this co-processor card seems to be out of production. I know there are a few on eBay.

Are there alternatives? How does this co-processor compare to simply using multiple CPUs? Thanks for your answer.

[AndyJP]

Not sure about Phi. But the code for CUDA works
1 - with simple isotropic dielectrics only. No anisotropy, or ferrite anisotropy is handled by GPU code.
2 - Non-enterprise GPUs are now artificially locked from execution. Even though Geforce supports all the CUDA capabilities of TESLA and Quadro, it was locked by some very dark(and forbidden even to guess) reasons from use.
3 - Only generations of nVidia Quadro and Tesla boards listed in the comptibility list are supported.
4 - Since ANSYS may lock non-listed Quadro and Tesla boards any time, despite the equivalence of CUDA capabilities with the listed boards, you have to consider purchasing exactly the boards from the list. Even if those are 10 times expensive and inferior to other boards on the market.... At least buy from the same series.
Not very nice, indeed. Particularly when comparing with the other software on the market.

[GeorgM]

Thanks Andy as far as I understand the direct solver cannot be accelerated with a GPU. This works in part for the iterative solver and reasonably well for the transient solver. According to the documentation, the direct solver can only be accelerated by the Phi, which is not a GPU. No clue about what is the difference to a GPU. The obvious difference is that you cannot use a MIC to drive a display.

[AndyJP]

>as far as I understand the direct solver cannot be accelerated with a GPU.
As I said, it can be accelerated with GPU, unless you use anisotropic media in your design. there may be more restrictions, which I did not find. ANSYS treacherously refuses to disclose all the limitations and requirements. But with simple copper on isotropic PCB the nVidia Quadro RTX board gives a well visible acceleration at the matrix solving step.
I was running it with Quadro used for the main display at the same time.

The restrictions on use of Geforce are purely artificial, probably defined by the management team.

[GeorgM]

Thanks again. I misunderstood that and I'll give it a try.