Jump to content
Rahul Ponginan

Hardware Setup best suited for FEM-Simulation

Recommended Posts

This Topic is to discuss following questions:

What Hardware is best suited for Hypermesh / Radioss work frame in order to produce fast results. Following questions specify this question:

a) What Processors should one use best. Which Socket is more efficient

B) Some thing on Motherboards

c) RAM: More is better, but which frequencies are supported for the solvers: 1600mhz and above interesting?

d) Graphic Cards. As it sais somewhere, the more consumer related Cards like GTX Series from Geforce are not fully supported in Hypermesh. Is this an issue?

e) Cooling Aspects: Does Water Cooling affect solving time? Over-clocking useful?

This all points somehow to the discussion: Build your own Workstation Computer for FEM - Analysis or is the consumer Computer suited enough for the task?

As a reference consumer Pc, the following data can serve:

Motherboard: Asus Republic of Gamers FORMULA V
Processor: i5-3750k; 3,4Ghz; Air Cooled; Socket: 1155
Ram: Corsaier 4x4Gb 1600Mhz
Hyperworks on SSD Drive

The aim of this post is to give a guideline on how to optimize the hardware aspects of the system doing the FEM-Analysis, now and as a glance into the future: for further versions of Hyperworks with even more multi core support; better memory usage etc.

Thanks for all, who contribute to this.

Best Regards,

Manuel

Share this post


Link to post
Share on other sites

Manuel,

Thanks for creating this important topic, below is from an earlier post on a similar topic, this is perhaps a good place to start the discussion

Motherboards and Front Side Bus (FSB) speeds and the RAM memory speeds and types supported are important since this will determine overall performance and bypass any bottlenecks you may face with applications.

A good solver machine would benefit from a 1066MHz/1333MHz FSB+DDR2/3 ram speeds (currently topping out at ~1600MHz) and top Intel/AMD chipsets and memory. Having the clock timings equal to CPU clock speeds will have better performance (no delay in CPU and memory cycles).

I would recommend at the very least 4GB of RAM memory with 8GB-32GB to be a better choice (4GB or 6GB per CPU is a rule of thumb) to be able to run applications in core (memory) instead of having the system utilize virtualization memory management and/or swap space memory which is much slower. Optistruct\Radioss\MotionSolve\etc. jobs will run better in core than on swap or virtual memory.

For graphics, if it is necessary (which solvers machines do not require since graphic intensive jobs arent utilized) Ì would recommend NVIDIA Quadro FX series cards or the ATI FireGL/FirePro series. (256MB or 512MB -> whichever is more affordable).

The recommended card should have 1GB or more dedicated memory onboard and at higher speeds (DDR4/5). These cards have a GPU running with workstation specific drivers for CAE/CAD specific applications which should handle the graphics in more of a efficient manner. The need for GPU and on board ram is also recommended

As for the hard drives, the bigger and higher rpm speeds the better. A SATA II RAID 0 (performance mode) with a swap file at least 1.5 times the size of system RAM (example, 4GB RAM = 6 GB swap file). This should give a high burst rate and will run Optistruct\Radioss and MotionSolve jobs better and general graphics information throughput faster as well. There is also the formatting of the drives in particular cases that will definitely influence performance, example, transferring of usually larger data chucks may result in better performance if the hard disk is formatted in larger data clusters than the usual smaller size (less I/O calls to transfer data).
There are newer Solid State Drives (SSD) with excellent throughput speeds for I/O but at the expense of cost and disk space size.

Share this post


Link to post
Share on other sites

Hi Rahul,

then I would suggest we start at the first topic you mentioned:

Motherboards

So there is the front side bus speed which is motherboard dependent. Also the socket is determined with the motherboard. The motherboards is pretty important. The Support of higher tacted RAM is also depending on the motherboard.

So we start with the specifications:

Socket/Chipset: I know there are certain Sockets where the Workstation CPUs guarantee a limited number on how many failure lines the CPU does within a certain time frame. These processors are pretty expensive and therefore it is of value to discuss the importance of such CPUs. Because the first criteria for our motherboard will be the socket, let's start with that:

a) Is there a CPU Type one could recommend over others specific for Hyperworks and its Sovlers. Why?

We can limit this to all quad (and +) core CPUs since no one would use a single or double core for a new system.

I leave it with that as it would get to complicated if we get too detailed too soon. The Roadmap I have in mind for this Topic looks like the following:

1) Choosing the right motherboard (+CPU)
2) Choosing the right RAM
3) Choosing the right VGA
4) Choosing the right Hard Drive

Any suggestions are appreciated.

Share this post


Link to post
Share on other sites

Manuel,


The information in my post above needs to be definitely updated and complimented with an organized roadmap as you describe.


The roadmap looks good to me icon_smile.gif


should we also think about the kind of CAE the person intends to do, as HyperWorks has so many solutions for CAE.


One person may intend to use only Solid Thinking and this will be graphics intensive, another person may prefer to do FEA alone i.e. using just the solvers, another person may mesh with batchmesher and therefore does not need good graphics even for meshing so I mean here that we need to follow this roadmap for each main CAE operation and also someone like you who will do all/most aspects of CAE.


and perhaps we should also think about the operating systems that work best for the kind of CAE the user would like to do ? Linux for someone doing FEA alone, Linux to use the processing capacity of the computer better, Windows is easy and good for SolidThikning or maybe MAC has its own advantages etc


Another thing to discuss is how are computers available in the market and if there is such flexibility to create a computer with different individual components, will everyone choose to do this kind of assembly? would people prefer to go for stock configurations more? also what factors need to be considered to get these "best for HyperWorks" individual components working together well i.e. get the motherboard and vga and the RAM etc integrate well ? what will be their combined working performance?


I suppose as we work towards understanding what each hardware component does generally, how it is used for CAE, FEA , what is available in the market and what will be available in the future we will eventually arrive at the best computer for FEA and the best for HyperWorks.


So I have to do some research and get back when I have good information icon_wink.gif as you say all suggestions by everyone is very much appreciated icon_smile.gif


Share this post


Link to post
Share on other sites

Yes,


I did not thought about the person sitting behind the desk. You got a good point there. An Idea that came up to me during the day: We could really use some data on how fast the solver works depending on the hardware. Imagine the following:


You guys at altair create a reference model, which uses the solving mechanisms we want to do our research on. For example, a nonlinear analysis. These models may even exist today at your end. These models could be then spread out here and as we let the solver run on the pc , we can send the data from the stack file, on how long the cpu was active in which area. So we could get some data and may check the influence of the parameters.


As I am pretty sure, that something like this already exists. Do you know more about it?


Share this post


Link to post
Share on other sites

Graphics

For graphics, if it is necessary (which solvers machines do not require since graphic intensive jobs arent utilized) we would recommend NVIDIA Quadro FX series cards or the ATI FireGL/FirePro series. The recommended card should have 1GB or more dedicated memory onboard and at higher speeds (DDR4/5). These cards have a GPU running with workstation specific drivers for CAE/CAD specific applications which should handle the graphics in more of a efficient manner. The need for GPU and on board ram is also recommended.

RAM

A Solver has no limitation in utilizing RAM - the faster the better,

A good solver machine would benefit from a 1066MHz/1333MHz FSB+DDR2/3 ram speeds (currently topping out at ~1600MHz) and top Intel/AMD chipsets and memory. Having the clock timings equal to CPU clock speeds will have better performance (no delay in CPU and memory cycles).

For RADIOSS or AcuSolve specifically the speed of RAM is the major factor. We would recommend at the very least 4GB of RAM memory with 8GB-32GB to be a better choice (4GB or 6GB per CPU is a rule of thumb)

For Optistruct one should aim for enough memory to satisfy RAM estimates printed by OptiStruct, preferably for in-core solver mode. If in-core requirements are too high to be practical (depending on expected size of typical run) it is still much better if OptiStruct run will fit in RAM instead of swapping out.

For Radioss and MotionSolve also in core solving should be achieved, rather than, on swap or virtual memory which are much slower.

Hard Drives-

RPM speed and size - The bigger size and higher rpm speeds the better.

Swap file - A SATA II RAID 0 (performance mode) with a swap file at least 1.5 times the size of system RAM (example, 4GB RAM = 6 GB swap file however if you have plenty of RAM, the rule that a swap file is 1.5 times the RAM no longer applies - swaps bigger than 20-30 GB are not practical.) This should give a high burst rate and will run Optistruct\Radioss and MotionSolve jobs better.

A SATA II RAID 0 (performance mode) with a swap file at least 1.5 times the size of system RAM will ensure general graphics information throughput faster as well.

For OptiStruct specifically, it is recommended to have dedicated separate scratch disk system as fast as possible (RAID 0, as noted above, but separated from everything else, no system areas, no swap, etc. - separate spindles, not separate partitions on disks otherwise used for anything).

Formatting of the drive - The formatting of the hard drives in particular cases will definitely influence performance, example, transferring of usually larger data chunks may result in better performance if the hard disk is formatted in larger data clusters than the usual smaller size (less I/O calls to transfer data). There are newer Solid State Drives (SSD) with excellent throughput speeds for I/O but at the expense of cost and disk space size.
But too large data clusters may actually slow down operations In one case with expensive RAID array of SSD's formatted with 1GB clusters - this gave tremendous speed in sequential I/O but choked solver or postprocessing in OptiStruct.
It is advised to format with large clusters in windows, On Linux however one is advised to format with non-journaling file system, this gives a better performance (e.g. xfs, instead of ext3, gave difference up to 20% for large jobs).

overclocking - Only advised for IT administrator with experience in overclocking, (as this is more appropriate for hobbyists) It can give speed boost (5-10%, seldom more), but it can make computer unstable, or even risk damage.

Rahul Ponginan likes this

Share this post


Link to post
Share on other sites

Rahul/Manuel,


Need assistance regarding hardware configuration for Hypermesh 13.0 for models with 20 to 25 million nodes and 5000 to 6000 contact/contact pairs. 


 


Rohit iyer


Share this post


Link to post
Share on other sites
Guest

Hi Rohit,


 


INDIA support is taking care of your request. Someone will get in touch with you from commercial support, soon.


Share this post


Link to post
Share on other sites
Guest

Rahul/Manuel,

Need assistance regarding hardware configuration for Hypermesh 13.0 for models with 20 to 25 million nodes and 5000 to 6000 contact/contact pairs. 

 

Rohit iyer

Could you share us your experience about that? I'm curious to know :)

Share this post


Link to post
Share on other sites

Me, too

I am just about to build a new system. Cant decide between the 5820k with the 2011 v3 socket and the 6700k with the 1151 socket.

What do you guys think about it?

Best regards,

Manuel

Share this post


Link to post
Share on other sites

Join the conversation

You can post now and register later. If you have an account, sign in now to post with your account.

Guest
You are posting as a guest. If you have an account, please sign in.
Reply to this topic...

×   Pasted as rich text.   Paste as plain text instead

  Only 75 emoji are allowed.

×   Your link has been automatically embedded.   Display as a link instead

×   Your previous content has been restored.   Clear editor

×   You cannot paste images directly. Upload or insert images from URL.


×
×
  • Create New...