[PedFr0]

Hello all

I am currently in need of a pc to use for FLUENT and would like some help on the issues regarding the RAM and processor capabilities:
1) If i were to use standard RANS turbulence models in double precision and had 16/32GB of ram, approximately how large a mesh could I solve?
2) Does the processor affect the size of problem that can be solved or merely the time it will take to solve, so for example, I could have 20million cells and solve on a dual core processor, but it would just take impractically long?
3) As far as the scale of the problem is concerned, would RAM or the processor be of more importance?

Thanks in advance

[CapSizer]

1. It depends. The coupled solvers need more memory than the segregated ones. Around 1 million cells / GB is a handy rule of thumb, but that is for single precision. Double precision coupled compressible could need up to 4 GB / million, but that is just a thumbsuck. I think hexa cells may need more memory than tetra (8 nodes vs. 4 per cell).
2. Correct, problem size is independent of CPU, provided that you have lots of time.
3. RAM.

[PedFr0]

I appreciate the response, and just in time as I can order my system tomorrow. I may just use single precision to allow for larger models. Thanks

[siefdi]

Hi All,

I have a sort question (again)?

Coming to the RAM, does ECC is that important in CFD?
I did some reasearch which succesfully confused me; some people says 'yes' while some other says 'no'.

Thanks a lot,
Regards,
siefdi

[CapSizer]

Originally Posted by siefdi

Hi All,
Coming to the RAM, does ECC is that important in CFD?

Check the motherboard specification. Dual-socket boards tend to require ECC, especially above a certain size. Single socket boards don't. In my experience of running CFD on single socket boards. it is not necessary to use ECC memory. I've seen no issues that could be attributed to not using ECC memory, but it would be interesting to hear what others have experienced.


There is an advantage to using normal RAM, in the sense that you can get faster memory (if your board supports it). It is possible to buy cheap, poor quality ECC modules, and these are to be avoided. If you buy quality brand name modules, you will be far less likely to have reliability issues. Brands that I know are satisfactory are Kingston and Crucial. There are other good ones out there as well.

[PedFr0]

While this thread is still receiving some attention, I would also like to get your opinion as to which is better, an entry level proffessional GPU, such as a quadro k600, or a mid-level gaming GPU, such as an Nvidia TX650Ti. I know that it only plays a role in the extraction of results, but less lag is always better.

[evcelica]

I don't think it plays a role in extraction of results. That's usually a hard drive operation.

[PedFr0]

In terms of contour plots and pathlines though, won't the GPU make a difference with regards to that?

[CapSizer]

I've had to ask myself this question a few times when putting together a new machine, and decided that for the kind of work that I was doing (< 15 million cells) I was better off spending the extra money on better or more memory, rather than on a "pro" graphics card. I'm not sure of that is right though. I've seen marketing material from Ansys showing screen images with obvious problems on a gamer's card, compared to the correct image from a certified pro card. Also, I understand that although some of the Ansys suite of programs can benefit performance-wise from the GPGPU, it has to be a Quadro rather than a GeForce.

I've also picked up difficulties with GUI instability when trying to use StarCCM+ with an ATI/AMD Radeon card, yet it worked fine with all my Nvidia GeForce cards. Perhaps it would have been OK with an ATI/AMD Fire Pro card? Dunno, didn't have one to test.

FWIW, I also have a compact "cheap" perpetually upgraded machine that has done much productive work for me, on grids of up to 12 million cells. Currently this just has the built-in Radeon graphics processor (no separate graphics card), and it still manages to do most of the required post-processing without too much difficulty.

I guess it partly depends on how demanding your graphics post-processing is likely to be, but it looks to me as if you only really need very heavyweight graphics equipment if you are dealing with larger data sets.

[PedFr0]

Thanks a ton CapSizer, your help has been extremely valuable to me, and from what I can tell, better to be safe and go with a pro graphics option.