[caitoc]

Hi,
I plan to DIY a workstation from origin PC to do CFD analysis.

Can you tell me Which hardwares are more important?
CPU, Memory...

Is a high performance graphic card important?

Thank you.

[CapSizer]

You should read previous postings in this forum, as this topic has been discussed at length several times in the last few months. To summarise it:

1. Get enough memory
2. Get fast memory
3. Get a system with many parallel memory channels (latest core-i7 systems are ideal, but cost quite a bit of money)
4. Get reliable brand-name memory
5. Find a motherboard that will support your super-fast memory
6. It is probably best to buy one or two notches down from the top of the range CPU. Speed in CFD is determined more by the memory than the CPU clock, and in any event you can normally overclock these CPU's.
7. It is worth paying more to get a good quality power supply, it will be quieter and possibly more reliable.
8. Install good cooling for the CPU & the case. Liquid CPU cooling tends to be quieter, and can cope better with the heat produced by sustained CPU computations.
9. Good graphics cards are always nice, but if you are building models that will run on a single socket homebuilt workstation, you probably don't need to prioritize this too much.

[ignatius]

In my opinion, a top range cpu is not only about clock, features like more cache, QPI or AVX are particularly useful when it comes to intensive floating point operations. I don't have any benchmark here but a Xeon will considerably outperform a i7, for the same clock and same RAM. I recommend you to do some research on this.

[caitoc]

Good. Thank you for you two's advice.

So here is my selection of configuration with an estimated cost:


Case: Enermax Fulmo GT 220
Motherboard: EVGA SRX 686
System Cooling: Intel liquid cooling 78.00
Processors: Dual Intel XEON E5-2650 2.0GHz LGA 2011 Octo-Core Processors (20MB L3 Cache) 1100X2
Power Supply: 1.5 Kilowatt Silverstone Strider 300
Graphic Cards: Single 1GB GDDR3 NVIDIA Quadro 2000 387
Memory: 96GB Kingston ECC Registered DDR3 1333MHz (12x8GB) 67x8
Operating System: Genuine MS Windows 7 Professional 64-Bit Edition 264
Hard Drive One: 1TB Western Digital Caviar Black SATA 6.0Gb/s, 7200RPM, 64MB Cache 90X2
Optical Drive One: LG DVD writer 25x2
Logitech mk710 mouse/keyboard 75
ASUS 24" ProArt Series PA246Q 450
--------------------------------------------------------------------------------

Total: (amazon price $5426)


Ignore things like monitor and keyboard. I use dual i-7 16-core processors and 96G memeoy with 1.5k power supply but only a 1G graphic card.

This workstation is used mainly to run commercial CFD software such as ANSYS, Fluent and CFX. My friend told me that graphic card will influence post-process if there will be plenty of gridding.

So I am not sure whether the graphic card will become the bottleneck of its performance. If so, I may reduce the CPU to 12 cores and 64G memory and get a 4G memory graphic card.

[caitoc]

I am actually a graduate student in school. I think it's worth for us to have high speed in postprocessing because we make a lot of report, slides and paper.

My colleague said that 16-core processors are much more imporant than a 2G graphics card and 96GB memory.

[evcelica]

Originally Posted by ignatius

In my opinion, a top range cpu is not only about clock, features like more cache, QPI or AVX are particularly useful when it comes to intensive floating point operations. I don't have any benchmark here but a Xeon will considerably outperform a i7, for the same clock and same RAM. I recommend you to do some research on this.

Wow, that's definately NOT correct. QPI allows two processors to talk to each other. AVX is a feature on sandy bridge i7s as well as Xeons. A XEON will definitely not outperform a good i7 SB-E.

Also If you are going to be running CFX you will be limited by memory bandwidth after 4 cores, so getting an octo-core CPU would be a waste of money. Two quad core XEON E5-2643 in a dual socket board, or better yet, two i7-3930K machines running in parallel would be a faster/cheaper option.

Check this link for more info: http://www.cfd-online.com/Forums/cfx...dware-cfx.html

I'd be happy run some CFX benchmarks with my two i7 machines against any $10K+ dual socket XEON you can find. I paid about $2K for each, so $4K total, and I'm sure they would blow away the XEONs no problem.

Also, get at least 1600MHz memory, not 1333MHz.

[caitoc]

Originally Posted by evcelica

Wow, that's definately NOT correct. QPI allows two processors to talk to each other. AVX is a feature on sandy bridge i7s as well as Xeons. A XEON will definitely not outperform a good i7 SB-E.

Also If you are going to be running CFX you will be limited by memory bandwidth after 4 cores, so getting an octo-core CPU would be a waste of money. Two quad core XEON E5-2643 in a dual socket board, or better yet, two i7-3930K machines running in parallel would be a faster/cheaper option.

Check this link for more info: http://www.cfd-online.com/Forums/cfx...dware-cfx.html

I'd be happy run some CFX benchmarks with my two i7 machines against any $10K+ dual socket XEON you can find. I paid about $2K for each, so $4K total, and I'm sure they would blow away the XEONs no problem.

Also, get at least 1600MHz memory, not 1333MHz.

Good to know. Thx for advice.

At least currently we are mainly running Fluent not CFX.
On the contrary, my collegue said his experence is that a CPU is more impotant than memory. So he asked me to use dual i-7 16 core processors while reducing number of memories to 64GB or less.
Now I am sure that graphic card is not important.
I will change the CPU to i-7 and get 1600MHz memory which means the memory has higher writing and reading speed.

I'd like to sum up here after reading the topics you recommended for me.
1 CFX is coupled and Fluent (which we are actually using more) and ANSYS are segrated.
2 Benchmark shows that per-core performance is more important than number of cores for CFX and ANSYS
3 Therefore, choose higher memory bandwidth and frequency cpu (memory badwidth is actually a parameter for processors)
4 for each core, I just need 2GB memoy so if I use 12 core, I will only need 32G memory.

[CapSizer]

Originally Posted by caitoc

1 CFX is coupled and Fluent (which we are actually using more) and ANSYS are segrated.
2 Benchmark shows that per-core performance is more important than number of cores for CFX and ANSYS
3 Therefore, choose higher memory bandwidth and frequency cpu (memory badwidth is actually a parameter for processors)
4 for each core, I just need 2GB memoy so if I use 12 core, I will only need 32G memory.

It's not quite as simple as that:

1. Fluent has both segregated and coupled pressure-based solvers.
2. Also look at how your licensing works, it may be better to have a smaller number of faster cores, but not necessarily.
3. The more cores you have for a given memory configuration, the less important the CPU frequency becomes.
4. Once again, that depends, but between 2 and 4 GB per core is what most of us are using. Some solvers and meshers are more memory hungry than others, and you need to look at what you will be doing most of the time, and what your peak requirement is likely to be. SnappyHexMesh, for example, may require much more memory to generate a given mesh than simpleFoam will need for the solution.

[caitoc]

Originally Posted by CapSizer

It's not quite as simple as that:

1. Fluent has both segregated and coupled pressure-based solvers.
2. Also look at how your licensing works, it may be better to have a smaller number of faster cores, but not necessarily.
3. The more cores you have for a given memory configuration, the less important the CPU frequency becomes.
4. Once again, that depends, but between 2 and 4 GB per core is what most of us are using. Some solvers and meshers are more memory hungry than others, and you need to look at what you will be doing most of the time, and what your peak requirement is likely to be. SnappyHexMesh, for example, may require much more memory to generate a given mesh than simpleFoam will need for the solution.

That sounds also reasonable.

I just have one more question:
One of my friends who once studied in a computing laboraty told me that a 4X16G memory will have much higher performance than a 8X8G memory because the former can have 16G memory running at one time whille the later configuration has only 8G memory running at each time. I doubt whether this is right because it sounds really not reasonable. Could anyone confirm my idea?

[evcelica]

You are correct, there will be essentially no difference between using 4x16GB and 8x8GB RAM modules. Each setup could use 64GB at one time, no idea what your friend is talking about.

Like Capsizer says you should check your licencing, and see how many HPC licences you have available, and build your computer around that.

Also don't just blindly go for i7 because of my comment, I was purely disputing the previous comment about "a XEON will considerably outperform an i7" There are some features you can only get with XEON, like running dual socket motherboards with two CPUs in one computer. I run two separate i7 machines in parallel but this may not be for everyone as it is a bit harder to setup and run, and doesn't have the convenience of a single workstation. Also with an i7 the maximum RAM you can get is 64GB per CPU, where you can get much more with a dual CPU XEON server board, but at a much higher price.

The amount of RAM you have will determine how large of a problem you can work on. The speed of your CPU and RAM (and number of cores) will determine how fast you can solve your problem.