CFD Online Discussion Forums

CFD Online Discussion Forums (
-   Hardware (
-   -   Help with build down under (

rpas1231 February 20, 2012 19:48

Help with build down under
I am planning to build a student desktop for my thesis running fluent/cfx (in windows 7) with about 5-10million cells/elements. No gaming just movies and want to overclock
budget of $1200 AUD

here is the list of:
  • IntelŪ Desktop Board DZ68BC ($256)
  • G.SKILLS 16GB KIT (4x4GB) DDR3-1866 ($138) or 12G Kit(4Gx3)2000 Patriot Viper2-Sector7 ($99)
  • CPU: Intel S2011 i7-3930K 6 Core ($649) or Intel CPUs Core i7 2700k ($343)
  • CPU fan: Noctua NH-D14 SE2011 LGA2011 ($86)
  • case: Thermaltake VN65021W2A SpaceCraft VF-I USB3.0 with 500W PSU($105)
  • cheap monitor, mouse keyboard, fans speakers ect ect ($166)

my question is the i7-3930K worth the extra cost for my simulations (difference $306)? Which of the two CPUs will be replaced first with a cheaper processor? And spending an extra ~30% in cost, will that increase my simulation speeds by 30%?

Lastly is 12GB of RAM sufficient for my simulations of 5-10 million cells (plus general use). And is the clock speed of the RAM enough?

Thank you,

edit added PSU info under case point

scipy February 22, 2012 16:55

1. Motherboard - wrong choice I'm afraid. You should go for a ASUS P9X79 (Deluxe or WS) or a MSI X79A-GD65 (8D). The reason is purely the 8 DIMM slots instead of the Intel's 4. This way you can get 32 GB RAM made up of 8 cheaper 4 GB DIMMs.

2. The RAM you picked is too fast and expensive for simulation needs. I've had Kingston HyperX CL9 (4x4 GB) 1600 MHz and then moved to a 32 GB G.Skill kit (4x8 GB) 1333 MHz and there was not even a 2 % decrease in the number of seconds per iteration. In any case, it's MUCH more important to go for more RAM instead of going for faster RAM (since the size of your mesh depends only on the amount of memory). In any case, LGA 2011 boards with 8 memory slots work with quad channel DDR3 which has been proven (in testing) to be even less dependent on memory speed for performance. So, either get 4x4 GB now, or just get 8x4 right away, you won't regret it. The reason I moved to 32 GB was the mesh size limit of ~8 million elements when solving a double precision problem with 16 GB RAM. Ideally, you'd buy 64 GB (8x8) and be done with it, however, people usually recommend not going over 2-4 GB per core because your cpu time cost becomes too great and it'd make more sense to build two computers with 2 CPUs and put 32 GB RAM in each. In this case, my 32 GB are also too much for the i7-2600K that I have, but since I'm a college student doing the thesis and there's no "time limit" pressure, I don't really mind waiting for a day or even more for a run to finish. Conclusion: get the 1600 MHz RAM since it's almost as cheap as the 1333, and get either 16 or 32 GB (consisting of 4 GB DIMMs).

3. I would've gone for the 3930K if the money situation would've allowed, but it didn't so I ended up with a i7-2600K. However, an FX-8150 would've been a better choice as it turned out. You can check in another thread I posted in the forum. But surely the 3930K will eat the FX-8150 since the cores are supposed to be around 20 % faster than 2600K's cores and you've got 6 (12) instead of 4 (8). This also answers your question about the % speed increase, 2 more cores mean 50 % more cores compared to a 2600/2700K and even if scaling isn't completely linear, you should be looking at a > 20 % overall increase in performance (probably even more).

4. The CPU cooler. If you're able to go with a top of the line air cooling like the Noctua, why not opt for at least a Corsair H80 water cooling system? It is in the same price range and the performance under normal loads (non-extreme overclocking) is within 2°C of the Corsair H100 (which I've got on my 2600K). I'd recommend this option since it's quieter, cooler and generally just better. :)

5. Case.. Doesn't matter much, but if you end up buying a Corsair H100 cooling, it'd be smart to make sure the case has a planned space for two 120 mm vents on it's top. For a similar price you can get the Corsair 400R or similar. However, it's best to ask the shop if the H100 fits inside the case with a 8 DIMM LGA 2011 board. Because of the power consumptions of those boards, they have large heatsinks on the chipseat above the left DIMM sockets which sometimes collides with a water cooling radiator in smaller cases like the 400R. But any case with room for 2x120 mm vents on top will do.

6. I don't think skimming on the monitor is the smart thing to do - it is the component that you look at the most. However, there's been nothing but good things said about the Dell U2312 and it's priced very competitively. About the mouse/keyboard.. get anything. :p

Now about the extra questions: I don't think either will be replaced. The 3930K is the most popular out of the 2011 socket processors since it's about half the price of the 3960X and it's the only other one with 6 cores. Similarly, the 2700K is the top end of the 1155 socket but.. there is the IvyBridge coming pretty soon (next year or so) but since Intel went back on their word that IvyBridge would fit into socket 1155 - it won't be replacing the i7-2700K sandy bridge. So, go for which ever one you can afford. Keep in mind that the LGA 2011 board is twice the price of a LGA 1155 but can easily fit 32 GB RAM (or 64 later on, if and when 8 GB DIMMs become more affordable).

Another recommendation that I'd make is to get an SSD hard drive for your system drive. The HDD is the bottleneck of performance in a modern PC. I can't tell you what kind of a transformation it was when I stuck the Corsair Force Series 3 120 GB SSD in there.. I've got Windows and ANSYS installed on it and keep all the "current" projects on it. When something is done and dusted, I move it to a regular magnetic drive for storage to keep room available on the SSD. One thing to mention here is: most of the modern SSD's suffer from a flaw that their speed decreases quite a bit when they're over 70 % filled, this is solved on Corsair Force Series 3 GT drives (Red ones) so get that instead of the regular one. Even a 60 GB SSD drive is quite enough, after all the important install files that are regularly used, there's still over 35 GB free for project/work files. You can install not frequently used programs like burning software and other crap on the magnetic drive. Things I keep on the SSD are things that get run several times in a day (Windows, ANSYS, Word/Excel, PDF/Acrobat, browser etc), while Roxio, Photoshop, Corel and similar things that are run from time to time are residing on the D: drive.

There, I wrote a novel for no reason. A simple list of stuff would've probably sufficed.. Anyway, hope it's useful. :D

Fugacity February 23, 2012 11:22

If you're looking for a cheaper solution, you might want to look into Amazon Web Services Elastic Cloud Compute (EC2). Since you're working in Windows, you can create an instance, log in w/remote desktop & install your programs and save it as an image.

Then you can close/start the image whenever you need the computer. You can start the image, RD in and load up Fluent and get it running and just log out for a few days and let it churn. You only pay when your instances are running, so it's rather cheap but it might take awhile to download the results.

minosdis February 23, 2012 12:11

With the cost of a X79 motherboard being a minimum of $250 and the 3930K being around $700, wouldn't it be most cost effective to just build two 2600k systems having 8 physical cores for ~700 each, then one 3930K system with only 6 physical cores? Understandably, the 3930K has 12 MB L3 cache, vs the 2600K's 8 MB. Do you think the 3930K's cores really run that much more efficiently, to make a 6 (12) core system faster than a 2x4 (16) core mini-cluster?

scipy February 23, 2012 12:25


Originally Posted by minosdis (Post 345972)
With the cost of a X79 motherboard being a minimum of $250 and the 3930K being around $700, wouldn't it be most cost effective to just build two 2600k systems having 8 physical cores for ~700 each, then one 3930K system with only 6 physical cores? Understandably, the 3930K has 12 MB L3 cache, vs the 2600K's 8 MB. Do you think the 3930K's cores really run that much more efficiently, to make a 6 (12) core system faster than a 2x4 (16) core mini-cluster?

I didn't have a chance to test something like this but there are several problems.. There is no way two computers can communicate as fast as hardware on the same motherboard (that's why a dual socket board and two 4 core xeons like E5620 are still probably a better solution than two 2600K PCs), you could get them to scale better if you used Infiniband cards to provide the interconnect with very low latency, but it's still not as efficient as a single hardware component. Plus, you have to factor in that for two 2600K configurations you need two motherboards, two power supply units (cases can be left out), two SSD drives etc. If the 3930K single core is 20 % more efficient than a sandybridge core, and if you have 50 % more of them and you avoid the interconnect protocol bottleneck.. I think the single 3930K is a better/cheaper solution than two i7-2600K's.

I'm looking for a future buy for myself and I was evaluating similar options.. Dual Xeon, dual opteron or multiple "regular" computers inteconnected. Conclusion was this: for 2500 euros I can build either a dual Opteron 6272 with 32 cores and 64 GB RAM (there can obviously be more, but I was trying to compare with 4 nodes of regular PCs) or 4 regular computers made up of FX 8150 AMD processor with 8 cores (making this the same 32 cores in total) and each node with 16 GB RAM making it 64 GB in total. Cheap used Mellanox infiniband cards can be had from ebay for 30-50 USD/EUR, however if you're gonna connect more than two nodes you also need an Infiniband switch and those are at least 300 EUR.

If you factor in the hassle of making the infiniband work, installing linux on the 4 nodes and all the related problems.. it's a much simpler and elegant solution to just invest in a proper workstation setup with a dual socket board and two processors.

Another question is always the licensing.. I've written something on that in another thread comparing the i7 2600K and FX 8150. Basically, if you're able to pay tens of thousands of dollars for HPC licensing, then you're probably better off with per-core faster hardware like X5600 series Xeons with 40 Gbit/s Infiniband cards etc. But if you're looking for an entry level setup and you don't plan on paying for HPC licensing.. then a whole "workstation" can be had for the price of a single X5690 Xeon :) and it's going to have more FLOPS than the mentioned Xeon.

dahicke February 23, 2012 13:18

I am looking at a similar situation using OpenFOAM on a linux box. I am looking at an AMD 8 core or a faster by 0.5 GHz 6 core processor. Any SWAGs or opinions on which may be faster?

All times are GMT -4. The time now is 20:59.