[Tensian]

Hey guys,

I am planning to buy a new computer. I am going to do CFD, mainly with FLUENT and OpenFOAM under Ubuntu systems.

I have thought something like this:

AMD FX Series FX-9590 4.7Ghz 8X
Asrock 990FX Extreme3
Corsair Cooling Hydro Series H80i GT
Kingston HyperX Fury Black DDR3 1866MHz 16GB 2x8GB CL10 (2) =32 Gb RAM
Seagate Desktop 7200.14 2TB SATA3 64MB
Samsung 850 Evo SSD Series 250GB SATA3
Cooler Master Silencio 452
Nox NX 750W
Sapphire R9 270 2GB GDDR5
LG GH24NSC0 Rewriter DVD 24x Black
Asus VS278Q 27" LED
Logitech Wireless Combo MK270


What do you thinkg about it?

I am gonna use it for aerodynamic problems, mesh count around 10 million cells, both steady and transient analysis.

I want to know your thoughts! I don't know much about hardware for CFD!

Best regards,

Tensian

[Daveo643]

You don't have enough RAM for 10 million cells, much less 1/10 of that, at least according to my experience using Fluent.

[evcelica]

That CPU only has 2 memory channels. For CFD you should be looking at CPUs with 4 memory channels, such as the Ivy-Bridge-E or Haswell-E based CPUs. The i7-5820K is the cheapest one with 6 cores I believe. I don't know what AMD has with 4 memory channels, but that's the main thing you want to look for.

[HyperNova]

i have amd FX-8350, it consume lots of electricity but no efficient! go for an intel like 4790K or 5820K, i repeat dont buy AMD !

[HyperNova]

http://www.cfd-online.com/Forums/har...es-please.html

[Tensian]

First of all thank you very much for your replies.

Daveo643: I thought that RAM was enough. I have performed my OpenFOAM simulations in computers with 32 GB and I found no problems...I don't remember where, but I have read that a "general" rule for CFD is that you need at least 1 GB for 1 million cells...

evcelica: Thanks for coming to help. I have read some of your post but it is still no clear to me the info related to memory channels of processor/number of cores/RAM kits.

I guess you suggest something like:

a)4 memory channels----you need 4*8 GB of RAM memory to optimize performance.

b)2 memory channels---you need 2*16 GB of RAM (but this is worse than the first distribution, am I right?)

But for example i7 you suggest has 6 cores and 12 threads, so, any bottleneck is going to be produced with option a)? Should I ran mpirun with number of processors equal to 6 or 12? I have no idea about this.

Will not be better to have 4 memory channels (4 * 8 GBs) and 4 CPU cores?Or there is not any bottleneck here?

AMD seems to provide more cores, but the dual channel memory will make useless this high amount of cores? If I switch to Intel I will have to do some changes in the plan I posted (motherboard, DDR4 memory...)

Does ddr4 memory a really advantage? I have read that the only advantage provided is related to electricity afford...

HyperNova: Oh, I didn't expect such a radical comment! have you tried Intel processors to compare performance and electricity consumption? By the way, has AMD any CPU with 4 memory channels as evcelica suggested?

The choice between AMD or Intel is getting me crazy. Many people suggest that the high amount of cores of FX series is a plus, but soon people fight with power needed/temperature issues/hydro cooling (I will buy one, it doesn't matter if I go for AMD or Intel)...Of course AMD is quite cheaper, but if you are going to find problems with memory bandwith, why to go for AMD?

Anyway if AMD has a 4 channel memory CPU all is going to change...or not?

I am anxious for your comments, I am reading the whole chat of "Hardware" and there is a wide range of opinions...scipy, kyle, wyldkat...if you are reading this, what do you suggest?

THANKS!!!

[Tensian]

mmmm...I have found this. What do you think about these workstations?

http://www.titancomputers.com/AMD-Wo...uters-s/26.htm

[HyperNova]

i have 2 systems :
1)intel 4770 - Asus Z87K - 16Gig 1600
2)AMD FX-8350 - Asus 990FX Pro - 16Gig 1866

The first system is about 25 percent stronger with consuming 70 percent energy.the point is memory bandwidth, chexkout these 2 photos, i have done that with AIDA64. Althogh my AMD has 1866Mhz ram but the speed in writing is much lower which is very vital in CFD.

[flotus1]

Dont Worry about the size of RAM, 32GB will be enough.
Daveo643 must be using some highly complex additional models if he needs 320GB of ram for only 10 million cells.
However, if you decide to get the I7-5820k, get a mainboard with 8 DIMM-slots. Then you can still upgrade to 64GB later.

You are right assuming that as the core count increases with a constant amount of memory channels, the memory bandwidth becomes a bottleneck. However most cases still scale well with a core count larger than the number of memory channels available.
So 6 physical cores with "only" 4 memory channels still make sense.
For companies that have to pay tens of thousands of dollars for their fluent licenses per year this becomes an issue. They should in fact use the fastest Xeon processors available with 4 cores and 4 memory channels each to maximize performance per license.
If license costs are lower, you can go for a larger amount of cores.

AMD has some Opteron processors with 4 memory channels. But just as the FX processors, they are simply outdated. AMD basically stopped development in the high-end sector a few years ago, I assume because they were no longer competitive.
So any "high-end" processor you buy from AMD was already outperformed when it was released a few years ago.
That is why it is not a good idea to buy AMD processors nowadays if performance matters.

Concerning your question on DDR4: we have seen here many times that CFD performance scales very well with the available memory bandwidth. The I7-5820k not only has a higher nominal bandwidth per channel due to higher memory frequencies, but is one of the cheapest options to get 4 memory channels.
If you absolutely cant afford it, one of the I7-4xxx is the next best option, but performance in CFD applications will be significantly lower.

[Tensian]

Thanks again for the advises. It seems that nobody is going to support AMD. I have understood that for cheap and fast cores AMD was the good solution. But as all of you said, Intel will be the best choice...

HyperNova: thanks for the test. If I finally go for Intel, I will upload some benchhmarks too. I have never thought that RAM memory and writing times were so important in execution times...Any other hint? For example related to graphic card. Saphire, FirePro, Quadro....

flotus1: Nice explanation, but if in most cases CPU scales well, the same should be true even for AMD, it's to said, if i7 5820 with 6 cores provide good performance with 4*8 GBs RAM, why AMD FX 9590 cannot do it?

I am almost convinced with i7 5820...

Last question: what's the difference between DDR3 anmd DDR4? Is saving power the only advantage (that is a good advantage anyway, saving power is the same as saving money or CFD will be beneficiary in any other sense?

It's hard to believe that the Opteron machines showed in the link will lost the battle with Intel i7. They look like really computational monsters!

Thanks again for all the advises!

[flotus1]

Originally Posted by Tensian

flotus1: Nice explanation, but if in most cases CPU scales well, the same should be true even for AMD, it's to said, if i7 5820 with 6 cores provide good performance with 4*8 GBs RAM, why AMD FX 9590 cannot do it?

For several reasons, most of them were already stated:
The FX 9590 only has two memory channels. You dont increase the meory bandwidth by using 4 DIMMs instead of two. So from the memory bandwidth point of view, the I7 wins by a factor of at least two. On the cache side, the I7 has 15MB of L3-chache while the FX only has 8MB.
And then again the FX 9590 is an old processor architecture that was not competitive even during the release a few years ago. That is why AMD decided to sell overclocked CPUs with an unreasonable power consumption. So even if CFD performance scaled linearly with the amount of cores for the AMD CPU (which I doubt for 8 cores and only two memory channels), its performance per core is so low that 8 cores still deliver a lower performance than the 6 cores of the I7.

For now, there is no real benefit in DDR4 Ram apart from power consumption. If you are not limited to ECC RAM, you can get the same amount of bandwidth with both of them.
Of course this will change in the future since DDR3 is capped out and DDR4 will allow for even higher amounts of RAM and higher frequencies.

Why would you judge the performance of a CPU from the look of the PC case? Benchmarks are a much more accurate way to measure CPU performance:
https://www.pugetsystems.com/labs/ar...d-Opteron-587/
https://www.cpubenchmark.net/singleThread.html
https://www.cpubenchmark.net/high_end_cpus.html

[Tensian]

Thanks flotus1 for this nice explanation and useful links!

I have read that usual CPU benchmarks are not the best way to compare computers for CFD behaviour, but this gives an idea about how powefurl Intel is compared to AMD.

I am not good at English so maybe I am not understanding something, but this machine:

http://www.titancomputers.com/Titan-...-Se-p/a200.htm

seems to be more or less in the price range of my configuration (even changing FX to i7 and incorporation a Quadro graphcis card).

So, will this workstation win the battle against i7 5820 machine, even with an older architecture, overcloking, power consming and so on?

I just want to learn the most, I have no doubt about comments here, because I am sure they are made by CFD experts with much more experience than me. HArdware world looks very interesting and also spending this amounts of money is something that force myself to get the best money/performance relationship

Thanks for all the info, this thread is giving me much more knowledge than I expected at first!

[flotus1]

I personally would not put my money on that setup before seeing some actual CFD related benchmarks of it.
Until then, we can only try to estimate the performance.
The total memory bandwidth will be only slightly lower than on the single-socket intel I7.
AMD decided to provide some benchmarks for the floating point throughput of some similar processors:
the AMD 8-Core processors only reach between 70%-80% of the performance of an old 6-core Xeon processor from 2012 with a very low clock speed (1.9 GHz).
Extrapolating these results and again keeping in mind that it is more efficient to feed 6 fast cores instead of 12 slow ones with about the same memory bandwidth available, I would still opt for the solution with a single I7-5820k.

[Tensian]

Hi flotus1,

Thanks your your sincerity. Maybe I could upload a test case when I have the computer in order to make some comparissons with somebody who uses AMD processors...

It is interesting the energy consumming point of view. If you work for example at university enterprise, you are not in charge to pay the energy receipt, but if you are going to make simulations at home, this is something you have to consider....

Some comments about graphic cards? What do ou think about Saphire? Are they compatible with Intel? I would like to make something with CUDA in the future...GPU looks promising.

Cheers!

Tensi

[HyperNova]

hi again,
because i wanted to do my simulations it home power that pc uses was important, as u mentioned if you want to use at university yes the power is free, so i am moving my AMD system to the room which university gave me ,
GPU accerators are very promising but i think it needs more times, right now Quadro K80 is the most powerful GPU for engineering works, but expensive 5000$ more or less,

[Tensian]

Final setup:

Intel Core i7-5820K 3.3Ghz
MSI X99S SLI Plus
Corsair Cooling Hydro Series H80i GT
G.Skill Ripjaws 4 Black DDR4 2400 PC4-19200 32GB 4x8GB
Seagate Desktop 7200.14 2TB SATA3 64MB
Samsung 850 Evo SSD Series 250GB SATA3
Cooler Master Silence 452
Nox NX 750W
PNY Quadro K620 2GB DDR3
LG GH24NSC0 Recording DVD 24x Negra
Asus VS278Q 27" LED
Logitech Wireless Combo MK270

This has been a headache for months...I hope this will best choice for this budget...

Thanks everyone who have helped!

[flotus1]

One last thing: G.Skill Ripjaws DDR4 2400 32GB 4x8GB seems to be available at more or less the same price of a 2666 MHz Kit. Thats 11 percent more performance for free.

[Tensian]

You are right flotus, difference is about 10 euros!

I guess I can go for it, anyway I am exceeding my budget (this costs about 1850 euros). Do you find it expensive? Just about 300 euros compared with the initial setup I posted.

[Tensian]

I have to make some tests but...what will be better? Run an OpenFOAM case with 6 processors or 12?

Is this fact different in OpenFOAM and FLUENT? I have read that hyperthreading does not improve performance...

i7 5820 has 6 physical cores, so I am not sure how to use parallelization with this two different codes...

[flotus1]

You have a few more options to save some money without sacrificing system performance.
A Crucial BX100 SSD will be just as good as the Samsung 850 Evo. And for a system-SSD, 120GB are enough in my opinion.
Drop the water cooling and get something like the BeQuiet Dark Rock 3 or the Noctua NH-14 instead. They are still capable of cooling a slightly overclocked I7 5820k at a low noise level.
I dont know how much that 750W power supply costs, but a 500W power supply from a quality brand is sufficient for your setup.
And then there is this rather expensive 27" monitor that does not increase the computational performance at all

Concerning Hyperthreading: It CAN increase performance for some CFD calculations.
But I would not expect much more than a 10-15% increase going from 6 cores with HT turned off to 12 cores with HT turned on in most cases.
If you dont have to buy 6 additional licenses to do that, you can in fact use Hyperthreading for CFD.
In the worst case however, performance can even be lower due to the additional communication overhead between 12 mesh partitions instead of 6.
It depends on the cases you are running, you will have to try it. Feel free to tell us how your tests went afterwards.