Running HFSS on a GPU

Is it possible to setup HFSS to use NVIDIA'S tesla card. If so, how would one do this and what kind of performance increases does it yield.

nathj72 said:

Is it possible to setup HFSS to use NVIDIA'S tesla card. If so, how would one do this and what kind of performance increases does it yield.

Click to expand...

To the best of my knowledge, there is no such facility. Agilent recently pointed out to me one of the advantages of EMPro over HFSS is that EMPro supports the use of GPUs, so giving parallel computing for no extra cost. I actually sent the Agilent email to Ansys, as I wanted to clarify a few other things, but Ansys never said "we support" GPUs or anything like that.

Dave

HFSS does not support GPU acceleration. As of Release 14, various other ANSYS products do support GPU acceleration, but not HFSS. The reasoning is fairly sound... GPU acceleration is well suited to iterative processes, ie FDTD. However in HFSS, being an FEM solver, it is required to invert a sparse matrix (or dense matrix as in IE solver) and the inversion algorithm is not well suited to GPU acceleration. Also, GPUs tend to be restricted to relatively small memory limits as opposed to DDR3-4 RAM. This is not to say it will never happen... but as of now no support for GPU acceleration.

As for EMPro... well... if a user enjoys paying to be Agilents Quality Assurance volunteers then go for it (tons of bugs) It is better than it used to be, but still slower, requires more RAM, and limited features as compared to HFSS.

Have Fun

tallface65 said:

As for EMPro... well... if a user enjoys paying to be Agilents Quality Assurance volunteers then go for it (tons of bugs) It is better than it used to be, but still slower, requires more RAM, and limited features as compared to HFSS.

Click to expand...

I guess you work for Ansys?

Who I work for is no more important than who you work for... however I have been accused of being biased toward HFSS as I have used it for years and all through university and it has worked well for me

Have Fun

tallface65 said:

Who I work for is no more important than who you work for...

Click to expand...

Well, it does start to have some importance if you sing the praises of one piece of software (HFSS) and say another is poor (EMPro). It's not unreasonable for someone to question in that case if you work for Ansys, as you may have what's generally called a conflict of interest. Wikipedia has rules about people editing pages on articles where they have a conflict of interest. To my knowledge there are no such rules on this forum.

Your contributions are greatly appreciated - irrespective of who you work for. I'd be interested in hearing more about a comparison of HFSS and EMPro from you if you have used both tools. But clearly such a comparison would be less objective if you have a conflict of interest. That's still not to say it would not be useful.

Clearly it is up to you whether you want to say who you work for, but I think you need to appreciate others would be interested in a simple Yes/No answer to whether you work for Ansys.

Dave

tallface65 said:

Who I work for is no more important than who you work for...

Click to expand...

That's not a secret - I post under my real name and work as a consultant and reseller for Sonnet.

tallface65 said:

however I have been accused of being biased toward HFSS as I have used it for years and all through university and it has worked well for me

Click to expand...

It's great if EDA experts support other users, no matter who they work for. Unfortunately it's easy to abuse these forums for propaganda.

As far as the initial question is concerned, I fully agree with what you wrote on GPU for FEM codes. It seems to be much more efficient for time domain codes.

One can look at all posts I have made and it will become obvious that I work deeply with HFSS daily in many different applications. However, as for conflicts of interest, I am not representing any company (my employer or ANSYS) when I assist users on this forum and thus I may state opinions formed from any previous experiences had. In the case of my EMPro statement, it was hearsay from 3 colleagues of whom I have befriended at other companies who tried the latest version of EMPro, so for full disclosure not based upon my own usage.

Also, as for propaganda, I only post helpful responses to questions that I can answer... I rarely speak poorly of other tools, as I have not used them in years, but if I hear of issues from colleagues that I trust, I have no hesitation to post them for perspective.

Have Fun

tallface65 said:

HFSS does not support GPU acceleration. As of Release 14, various other ANSYS products do support GPU acceleration, but not HFSS. The reasoning is fairly sound... GPU acceleration is well suited to iterative processes, ie FDTD. However in HFSS, being an FEM solver, it is required to invert a sparse matrix (or dense matrix as in IE solver) and the inversion algorithm is not well suited to GPU acceleration.

Click to expand...

But HFSS does have a time-domain solver too. It has all three of the major methods for analysis of 3D EM structures - FEM, MoM and time-domain.

tallface65 said:

As for EMPro... well... if a user enjoys paying to be Agilents Quality Assurance volunteers then go for it (tons of bugs) It is better than it used to be, but still slower, requires more RAM, and limited features as compared to HFSS.

Click to expand...

I noticed a fair number of bugs in HFSS while I had it on trial. I could post a number of screen shots of where things went badly wrong. Some of them could be put down to the fact the Windows trial was for a laptop which did hibernate some times, but I'm not convinced that was the full issue. In any case, other programs seem to be able to work on laptops which hibernate.

As I commented in another thread, I've never seen anything sensible written about installing HFSS on Redhat 5, despite it is supposed to be supported.

Dave

It'd be interesting to look at the screenshots.

drkirkby said:

But HFSS does have a time-domain solver too. It has all three of the major methods for analysis of 3D EM structures - FEM, MoM and time-domain.



I noticed a fair number of bugs in HFSS while I had it on trial. I could post a number of screen shots of where things went badly wrong. Some of them could be put down to the fact the Windows trial was for a laptop which did hibernate some times, but I'm not convinced that was the full issue. In any case, other programs seem to be able to work on laptops which hibernate.

As I commented in another thread, I've never seen anything sensible written about installing HFSS on Redhat 5, despite it is supposed to be supported.

Dave

Click to expand...

Yep, the problem with GPUs is that they can have 8 Gb of onboard memory on every card, but in most cases they are mirrored. So, with 4x8Gb you still get 8 Gb + synchronization penalty on every matrix manipulation.
And as you can see, fem solver uses no more than 4 cores in HFSS V.13. I can say, this is waste of resources that can be spent for another sweep step for example... but that's actually as simple as it is - you don't need much multithreading for FEM that GPU provides. But you need very high CPU clocks that Intel i7 provide. (Xeons are also useless here)

The only thing I am angry at is that HFSS 13 and 14 still don't use Intel AVX extensions, that could grant up to 200% performance in matrix problems.