[husker]

Hi,

I would like to know the experts' opinions about this CPU and its capability on CFD implementation.

It has 4 cores and 8 threads. The clock speed 3.06 GHz.

If there is enough RAM, how many tetrahedral and polyhedral elements can be resolved with incompressible segregated solver and compressible coupled solver?

Regards

[scipy]

If you haven't already, don't buy this CPU/system. I'll come back to why, but first let's answer your questions.

This CPU can adress 24 GB of RAM at maximum, this would mean that you could solve meshes up to ~15 million tetra elements in the segregated solver and probably up to ~10 million with a coupled solved. For double precision reduce these numbers by ~20 % or so. As far as polyhedrals go, you can solve a lower number of cells because elements as such have more nodes and take up more memory than tetras, but you also need less of them to acquire the same accuracy. However, if you were planning on producing tetra meshes and then doing a polyhedral conversion inside the solver (like in Fluent), the conversion process itself uses more memory than the subsequent solve of the polyhedral mesh - something to keep in mind (there needs to be overhead).

Now to reasons why not to buy it.

1. It's an outdated architecture (45 nm litography, 130 W TDP). What this means is that it's going to be slower AND use more power and require more cooling than a comparable CPU of a newer generation.
2. It does have support for triple channel memory, but it also supports only up to 1066 MHz DDR3 - this only goes for server/workstation motherboards with ECC support and locked BIOS features. If you were to go for a desktop board (X58 chipset, for example ASUS Sabertooth X58), then you could use memory of up to 1866 MHz. However, if you are going for a desktop board - you do not need a Xeon CPU (even though this CPU would most likely work with this board), because there are regular desktop version cpus like the i7/i7-Extreme which would be significantly cheaper and more suited to your applications. Another point since we're on the memory front: newer i5/i7's of IvyBridge architecture have the same memory bandwith with dual channel memory as this cpu has with tripple channel and are much more efficient per core (both performance and power wise).
3. Some real world results: I've recently started gathering some CFD benchmark results for the purpose of starting a thread here in which people could get some decent information on what's the best hardware combo for the price. Even though it's still in the early stages, I have enough information to provide you with some insight.

An i7-970 desktop CPU (made in 32 nm litography - meaning even more power efficient than yours, with better performance too) also has tripple channel memory and the same maximum bandwith. I've tested this CPU with my benchmark case (which is incompressible flow with a coupled solver in Fluent, ~3 million cells in the mesh consisting of prisms, tetras and hexas) which is made to fit inside 8 GB memory (uses around 5.1 GB in memory). I'll add a link to the full results a bit later, but for now here's what's important.. At 4 cores (which is what W3550 is limited to), the "better" i7-970 had the performance of 15,25 seconds per iteration (or 236 iterations/hour, however you want to look at it), while my "cheap" but new i5-3570K took around 14,2 sec/iter (253 iter/hour) at the same memory speed of 1333 MHz (with 9-9-9-24 timings, same as the ones on the i7-970).

The i5 can also address 32 GB of RAM and it supports memory speeds up to 2800 MHz. I currently have 32 GB (4x8 GB) of 1866 MHz (9-10-9-27) DDR3 and at those settings I get 11,18 sec/iter (322 iter/hour) performance. The TDP (thermal design power) of the i5 is 77 W, while your Xeon (for the same number of cores) is 130 W. Also, there is no need to buy processors with HyperThreading capabilities for CFD as most manufacturers' first recommended step is to turn that off as the software uses only physical cores. Previously I had an i7-2600K with 32 GB of 1333 MHz DDR3 and tested HT on/off to find that performance was similar (0.3 % slower with HyperThreading on) - so I sold the much more expensive i7 and the old RAM for a newer (more power efficient and with a faster memory controller) CPU and faster RAM. The current configuration is 26,5 % faster than the old one and scales nearly 10 % better on 4 cores than the i7 did (all because of the RAM, which is a huge bottleneck in CFD applications).

My recommendation: forget about Xeons (especially those such as W3550 which can only work in a single CPU configuration), and buy yourself a nice IvyBridge i5 such as i5-3550 or i5-3570K (K is the unlocked version that supports overclocking, but it's worth mentioning that for a potential 30 % increase in frequency you get only 6,3 % increase in performance which is not worth it), a decent Z77 chipset motherboard and as fast a RAM as you can afford. I'll post this too, but for now I'd just add that 1866 MHz DDR3 CL9 costs about 40 % more than the baseline 1333 MHz and you get around 25 % bump in performance, while a 2133 MHz CL9 costs around 110 % more than the 1333 MHz kit and you only get a 33 % bump in performance. Meaning that 1866 CL9 is the "best bang for buck" for now.

One more thing: CFD software makers recommend around 2-4 GB of RAM per CPU core, and I somewhat agree with this. This is what happens: if you get 32 GB of RAM and a single 4 core cpu, if you are going to use all the RAM up with your mesh/case - then your calculation (even with a coupled solver which converges faster) for a steady state case is going to take ~24 hours, which is a bit long. I'm a student and my PC was built with budgetary restrictions and for the purpose of solving my final thesis project cases for which I had plenty of time, so even if I had to wait a day - it wasn't such a big deal, as long as I could get as much detail in the mesh as possible. However, if you are buying this PC for "work" or something similar which comes with stricter deadlines - then going for that much RAM is not really financially responsible. If I could have my way I'd buy 4 of these i5's, each with 8 GB of RAM and network them all together to get a nice "mini-cluster" with the capabilities to solve 32 GB cases, but with 16 cores and not 4. So, an optimum would probably be 16 GB of memory - you'd use up to 12 or so, and still have headroom for polyhedral conversions etc..

If you have any questions, do ask.

[husker]

Hi,

Thank you very much.
You have given very much information in your answer.
I will keep your answer for reference because it includes both computer technics and CFD issues.

On the other hand, my budget is also limited so I will try to minimize the price. Then, I will try to build a system based on i5-3550 CPU, Z77 chipset motherboard and with 12GB 1866 MHz RAM.

What do you think about the graphics card? I just need to import, transform and visualize the CAD and Mesh model in a reasonable speed.

Kind Regards

[scipy]

You can't really get 12 GB of RAM (well, you can - but don't). The motherboard has 4 DIMM slots and for dual channel to work properly you have to either populate only 2 of them, or all 4 of them - never 1 or 3. So, you will get a 16 GB kit of 4x4 GB.

Any of these kits will do, however I would go for Corsair (G.Skill, at least according to my information gets a large % of their memory kits returned for failures - they do offer warranty and will replace it without question, but it can be a hassle).

http://geizhals.de/?cat=ramddr3&xf=2...06_1.50#xf_top

As far as graphics go, I don't know if we have the same definition of "reasonable speed". For example, for a long time I was using the onboard graphics that came with the i7-2600K (HD3000) and then i5's newer HD4000 onboard graphics, but none of these are made for OpenGL graphics which most CAD and mesh softwares use. For a while I had a GeForce GTX 560 Ti which had barely any improvement on the CAD/Mesh packages. I used SpecViewPERF to benchmark this (it uses a mix of catia, maya, proe, solidworks, siemens NX tests) and all regular Direct3D graphics cards that are made for gaming are going to have similar results. My 560 Ti had 6.1 fps in catia test, 6.06 in maya, 1.41 in proe and 10.1 in solidworks (which were the ones that mattered to me). On the other hand, my current AMD (ATi) FirePro V5800 has around 50 fps in catia, similar in maya, 40 in proe and most importantly for me 115+ fps in solidworks - where I do most my CAD stuff.

Also, mesh rendering performance in ICEM CFD has jumped hugely (I didn't know how to measure this), but before the V5800 if I rotated the surface mesh consisting of about 1 mil triangles it was very jittery and had maybe 1-2 fps (maybe even a frame every 2 seconds here and there, depending on the level of zoom). Now, with a FirePro V5800 I can start rotating the mesh MADLY by pulling the mouse left-right as fast as I can, and it renders it completely smoothly. It really increased the workflow speed, especially when you need to zoom in, out, rotate something, pan the mesh to do visual checks for bad elements etc. So, my recommendation is: get any professional OpenGL card that you can afford. Even a FirePro V3900 or V4900 is going to do a MUCH better job than any "gaming card" of up to 2x higher price tag.

Another thing: it seems that at the moment AMD/ATi is outperforming Nvidia when it comes to professional OpenGL graphics cards (take a look at this video: http://www.youtube.com/watch?v=a4ZMZwZqz9Y), and same should hold true when comparing lower priced models such as V5900 vs a Quadro 2000. Since AMD/ATi seems to be priced slightly more competitively - it makes it (for now) a better option.

Price comparison V5900 vs Q2000: http://geizhals.de/646502 vs http://geizhals.de/590095, you can find the V4900 and V3900 on the same site by doing a search. V3900 is in the ~100 euros range while V4900 is in the ~150 euros range. If you are not going to do large assemblies in CAD or huge multibody meshes, any one of these two will be more than fine. In fact, in regard to your budgetary restrictions and the rest of the PC - your only sane choices are V3900 or a V4900.

[scipy]

In fact, here's a configuration I'd recommend in the end:

Motherboard: ASRock Z77 Extreme4-M (http://geizhals.de/746841)
CPU: Intel Core i5-3570K (http://geizhals.de/761856)
RAM: Corsair 16 GB (4x4) 1866 MHz CL9 (http://geizhals.de/809395)
GPU: AMD FirePro 3D V4900 (http://geizhals.de/705597)

All in all ~540 euros, which is really not much for something so capable in engineering and CFD applications. In fact, this machine will outperform any DELL/HP workstation of up to 1000 eur, you can bet on it.

Even though overclocking doesn't get you anything in most of the CFD applications, the unlocked 3570K is only 10 % more expensive than the 3550, and will be much easier to sell at some point in the future (because they are in much higher demand by the gaming/overclocking community). Same goes for the motherboard (you could get a cheaper Pro3 version, but this one will retain more value and demand), if you think you need the 2 regular PCI slots (which are useless as far as I'm concerned because anything that used to go into the PCI is now onboard, such as LAN cards, sound cards and extra USB ports) then you can go for the full ATX format Extreme4 board (not -M). Graphics card - V4900 does have the DDR5 onboard (V3900 is really entry level and has the old DDR3) and is probably the "best buy" at this point.

As far as power supplies and cases go, any 400-500 W power supply will do just fine, so I'd base my choice on warranty lenght.. get any decent power supply in that power range which has 5 years warranty.

P.S. If you have another ~100 euros to spare, I'd highly recommend getting an SSD hard drive (OCZ Vertex4 128 GB model is the optimum for now, 5 years of warranty and great performance figures http://geizhals.de/759195). The "old" magnetic hard drives are the single largest bottleneck in modern pc performance. SSD would decrease your boot times, any applications (that are installed on it) start times and HUGELY reduce input/output time in CFD and FEM applications (especially if you do anything transient which has to do a lot of writes of case/data files every few iterations).

[husker]

You gave me much information than I've ever expect, thank you.

I will build my system with these components.

Kind Regards

[scipy]

Keep in mind, these should not be followed blindly.. If you are at a store or ordering online and there's a better deal on an ASUS Z77 motherboard, you should go for it. None of this is set in stone.

Also, I just checked gh.de for 4x4 GB 2133 MHz kits and 2400 MHz kits. It seems that the prices aren't that much higher. I was always looking for 4x8 GB kits and 8 GB DIMMs are always quite a bit more expensive than 4 GB, plus they're harder to get to work stably at higher frequencies.

For example, you've got these options too:

2133 MHz CL9 kits: http://geizhals.de/?cat=ramddr3&xf=2...~256_4x#xf_top

2400 MHz CL10 kits (9s are too expensive): http://geizhals.de/?cat=ramddr3&xf=2...~256_4x#xf_top

So, you could get a 2133 MHz kit that's only 15 euros more expensive than the 1866 CL9 kit, and it would be quite a bit faster for CFD (and increase your scaling efficiency when going from 1 to 4 cores).

Same goes for 2400 CL10 kits, but these are already 30 eur more expensive than the 2133 and since they're CL10 (one step slower cas latency timing) they wouldn't be much faster than the 2133 CL9 - making them not worth it. However, since 4x4 kits are usually packed like that because they're ought to support quad channel memory controllers used on X79 chipsets and Socket LGA2011, they are often more expensive than the comparable 2x4 GB kits (or any kits consisting of only 2 DIMMs which are aimed at dual channel boards), so in fact you can get two of these kits: http://geizhals.de/819307 which are 2400 MHz CL10. They will be marginally faster than the 2133 CL9, at a pretty low increase in price (120-130 euros for 2400 CL10 vs 110-115 euros for 2133 CL9)

So, in fact I'd change my RAM recommendation to at least 2133 CL9 if we're talking about 4 GB DIMM sizes (or 2400 CL10 if you can get two kits of 2x4 GB at a decent price wherever you're located).

[evcelica]

Good advice Scipy, definately one of the best bang for your buck configurations you can build. A build using an i7-3820 or i7-3930K would be better of course, but at a higher cost. You might want to think about going with the 3820 if you have a little extra money. If you want to compare apples to apples, for around $100 USD more you could get the i7-3820 with a four memory slot motherboard, use the same RAM, and have twice the memory bandwidth as the ivy-bridge (since SB-E uses quad channel memory, where ivy-bridge is only dual channel). If you have about $200 more you can get a motherboard with 8 RAM sockets and have double the RAM capacity, and double the memory bandwidth. You would then have to purchase the extra RAM as well for another ~$130 or so. Whether or not that gain is worth it is up to you though.

The one thing I would add to Scipy's advice would be to get a single RAM kit, not 2 kits. Mixing two sets of RAM kits, even identical ones, can sometimes give you problems. Especially since the XMP 1.3 memory profile might only be for two modules if you get a 2 module kit, so you could get frequent crashes until you figure out what everything is supposed to be set at when running 4 modules, which can be very difficult.
I've mixed two sets before and it worked perfectly; on another set of two kits it gave me a ton of problems, even when I set everything to what the exact same four module kit was set at. Its definitely worth the little bit of extra money to just get a 4 module kit.

[-mAx-]

Interesting thread since I am also in benchmarking phase and Intel W3550 is in part of it.
We had an old cluster (up to 14 desktops with 1Gb/s LAN) essentially made of P4 3.2-3 & 2.8GHz, and I ran some test cases from 1 to 10 million cells.
Colleagues also ran some cases on their computer, and I can give you some results
Most of involved desktop are from cad colleague (standard desktop: last year we received Intel Xeon W3550, and now we receive Intel Xeon E5-1620 )
-Test cases ran on OpenFOAM
-Involved CPU: *Intel Xeon E5620 QuadCore 2.4GHz (Desktop)
*Intel Xeon E5345 QuadCore 2.33GHz (Desktop)
*Intel Xeon W3520 QuadCore 2.67GHz (Desktop)
*Intel Xeon E5-1620 Quad Core 3.6GHz (Desktop)
*Intel Xeon W3550 QuadCore 3.07GHz (Desktop)
*Intel Core i5 3.2GHz (Desktop)
*AMD SixCore Opteron 8439SE 2.8GHz (Server WMWare)
*Intel Core i3 3.06GHz (personal iMac - VirtualBox)

Cores involved: 4
Iterations 400
1Million Cells

Intel Xeon E5-1620 Quad Core 3.6GHz (Desktop)----------1110 (s)
Intel Xeon E5620 QuadCore 2.4GHz (Desktop)-------------1822 (s)
Intel Xeon W3550 QuadCore 3.07GHz (Desktop)----------1856 (s)
Intel Xeon W3520 QuadCore 2.67GHz (Desktop)-----------2159 (s)
Intel Core i5 3.2GHz (Desktop)---------------------------2879 (s)
Intel Core i3 3.06GHz (personal iMac - VirtualBox)------2965 (s)
AMD SixCore Opteron 8439SE 2.8GHz (Server WMWare)---3253 (s)
Cluster--------------------------------------------------3591 (s)
Intel Xeon E5345 QuadCore 2.33GHz (Desktop)------------5817 (s)




Cores involved: 4
Iterations 400
2Million Cells

Intel Xeon E5-1620 Quad Core 3.6GHz (Desktop)----------2124 (s)
Intel Xeon E5620 QuadCore 2.4GHz (Desktop)-------------2995 (s)
Intel Xeon W3550 QuadCore 3.07GHz (Desktop)----------3530 (s)
Intel Xeon W3520 QuadCore 2.67GHz (Desktop)-----------4181 (s)
Intel Core i3 3.06GHz (personal iMac - VirtualBox)------5530 (s)
AMD SixCore Opteron 8439SE 2.8GHz (Server WMWare)---6407 (s)
Cluster--------------------------------------------------6615 (s)
Intel Xeon E5345 QuadCore 2.33GHz (Desktop)------------11430 (s)


I cannot explain why the AMD SixCore Opteron is so slow (Proliant DL785 G6)... But it is not the subject here