CPU and Graphics card for electrically large problems

Altair Forum User

Hi,

 

I have an electrically large model to solve, which I want to solve using the traditional MoM solution.

 

If I upgrade my machine to a CPU to 32 GB, with an 8 GB graphics card, how much of an improvement will I get in the processing time? Has there been a study done on the tradeoff?

 

Thanks

Mel

At the moment FEKO only supports a single GPU for the MoM and only for the LU decomposition phase.

 

You would need enough RAM to fit the model into.

 

You could also switch to higher order basis functions for the MoM if you don't have enough RAM for the default basis functions used.

Altair Forum User

Altair Forum User said:

You could also switch to higher order basis functions for the MoM if you don't have enough RAM for the default basis functions used.

Hi Mel,

 

I did this. In the solver settings, I checked on the Solve MoM with higher order basis functions (HOBF). Element order was set to Auto, range selection was set to Normal.

 

This greatly reduced the number of unknowns for my electrically large VHF model. I expected the simulation to go relatively quickly, but it is progressing painfully slowly. 

 

Wondering if there is any other setting I have to use in conjunction?

Mel

HOBF, in general, saves memory but not necessarily runtime.

 

Could you attach your model. Perhaps I can then make more model specific suggestions.

Altair Forum User

Due to the nature of my work, I'm unable to attach the model. I'll list the key differences.

 

Prior to using HOBF:

Used a standard FEKO mesh, resulting in ~120K unknowns.

 

Upon using HOBF and setting the Element Order to Auto, range selection to Normal, the unknowns came down to ~11K.

 

I'm simulating at 120-140 MHz at discrete intervals (8 in total).

 

These are the only changes. But the simulation is running incredibly slowly... 

Altair Forum User

Other than HOBF, is there any other method which reduces the total number of unknowns in the model??

Mel

You mentioned that you have 32 GB of RAM available. The standard mesh uses about a size of lam/12.

You need to reduce the number of unknowns to 60 000 to fit into 32 GB (will use about 28 GB).

This equates to increasing the mesh size by a factor of 1/sqrt(60000/120000), or lam/8.5 

This should still give reasonably good results.

 

Else, if you have a fast hard drive, like a SSD, or even a striped M.2, to use for a scratch file, the out-of-core MoM solver with standard basis functions should be faster than the HOBF.

 

Also if you can apply electric or magnetic symmetry, this will half the memory.

 

 

Altair Forum User

I think 32GB Memory would be enough but need a good workstation graphics card with fast 8GB memory.