Accelarator boards for EM simulation tools

[irfan1]

site:www.edaboard.com acceleware

I know that SEMCAD provides a hardware accelarator board for its simulation software. Does anyone know similar boards for instance for CST? I have heard that a company just released such a board with IBM Cell processor for scientific and such calculations. **broken link removed** The board is capable of an amazing 180GFLOPS of performance.

 

[doug freeman]

the semcad board comes from acceleware.com and my advice is to avoid as all they can do is accelerate a standard FDTD code, and only seem to increase speed noticably for certain situations (no open boundaries/dielectrics)

Accelerating CSTs FIT code is a much harder prospect (PBA etc.) and my feelings are that CST would not be too keen to become hardware vendors and all that this entails. I beleive they are looking at this for the future though, but let's see.

 

[irfan1]

interestingly, I beleive that CST has plans to employ hardware accelarators.

 

[turbocool]

I've used Semcad with acceleration and saw significant speed up for simulations in all cases including PML and dielectrics. Seems to me that the entire Semcad feature set is accelerated by the hardware.

 

[wave-maniac]

Other than, specific acceleration hardware, some academic research centers are trying to use graphic processors and boards, found in PCs, to expedite FDTD and similar simulation methods. The following paper was presented at IEEE MTTS 2006:

Accelerated Implementation of the S-MRTD Technique Using Graphics Processor Units
G. S. Baron, E. Fiume, C. D. Sarris, University of Toronto, Toronto, Canada
A time-domain electromagnetic modeling technique, namely a high-order scaling function based MRTD (S-MRTD), can be dramatically accelerated, through its implementation on commodity graphics hardware. This implementation is achieved by mapping the numerical operations of S-MRTD to graphics operations of a Graphic Processing Unit (GPU). The sustained speed-ups achieved, for two-dimensional problems, reach a factor of 30, significantly higher than any speed-ups reported for FDTD so far.

Here is the relevant website:

https://www.waves.utoronto.ca/prof/sarris/index.html

and here is the link to an overview of their recent projects (Pages 54-65):

https://www.waves.utoronto.ca/prof/sarris/Technical/Presentations/overview.pdf

 

[jkivi]

Hi all,

The Acceleware hardware supports 2nd order Higdon and CPML boundaries with material extrusion. The hardware also supports dielectric, magnetic and dispersive dielectric materials throughout the mesh. All of these features are accelerated by the hardware.

I currently work for Acceleware so if you have any questions feel free to pm me or check our website (https://www.acceleware.com).

-jeremy

 

[irfan1]

Would you be so kind to provide us a list of supported programs? For instance, Some of my collagues and myself wonder whether CST MWS and Rsoft Fullwave are supported by the accelarator board? Also I do not see any information regarding the price

 

[jkivi]

irfan1,

Our hardware acceleration of FDTD has been fully integrated by a number of partner companies. These companies set the price of hardware acceleration interfaced with their existing FDTD package. In order to receive a quote for a hardware accelerator, you need to talk directly with our partner companies.

If anyone is interested in a listing of companies we work with please feel free to contact me via pm.

-jeremy

 

[Mika3]

I´m intrested in the hardware acceleration too, but the acceleration cards always have 512MB or 1 GB Working Memory. what happens when I have a calculation, which needs 8 or 16 GB RAM. Can the acceleration card speed up these calculation really or only a little bit?

A little calculation example:
If i have a calculation which needs 16GB and I have a acceleration card with 1 GB. The card calculates 10 times faster than the PC-RAM. Without acceleration the calulation takes 100 hours.

My calculation with card:

cs ..... calculationspeed of the normal RAM
10cs ..calculationspeed of the normal RAM

(cs * 15GB + 10cs * 1GB) / 16GB = 25/16cs = 1,56cs

This means my calculation takes 100/1,56 hours = 64 hours

Is this calculation right, has anyone any experience with this ?

 

[jkivi]

Mika3,

A single Accelerator 400 card will fully accelerate simulations up to 28 Mcells in size. The ClusterInABox 800 will fully accelerate simulations up to 56 Mcells.

For simulations larger than the above, the Accelerator shares memory with the host. Assuming the speed-up factor used, your calculations seem reasonable.

Most of the simulations we have encountered can be fully accelerated (i.e. completely fit on the card(s)).

-jeremy

(1 Mcell = 1 million cells)

 

[wave-maniac]

This bit of news is interesting:
-----------------------------------------------------
CST and Acceleware announce hardware acceleration for CST MWS
----------------------------------------------------
Here is the link:

https://www.cst.com/Content/News/Details.aspx?newsId=92