[pagfetter]

Hello CFD Online!

This is my first post here and I'd like to start thanking you guys for the amount of useful information you make available in all these threads. They've been really helpful in my research during the last few weeks.

So, I work in a small engineering company and we're doing OpenFOAM analysis of race-car bodies. Our current cases involve symmetrical (half car) tetrahedral meshes with prism layer consisting of around 10 million elements. We use the STT k-omega turbulence model, double precision and a second-order coupled solver.

Our current machine is a makeshift solution with the following specs:

• 1x Intel i7 2600K (quad-core, 3.4 GHz, 8 MB cache, 21 GB/s mem. bandwidth, dual channel)
• 4x 8 GB DDR3 Kingston (1333 MHz, non-ECC)
• Gygabite GA-P67A-UD3-B3 motherboard

With it, we've been able to run our cases with almost full RAM and 100% on all 4 cores in about 5 hours. Time is not really a problem, since our modifications normally take days between each run, but mesh size is. For assymetrical (full car) or transient cases, we'll need much more RAM. That's why we're thinking of 128 GB RAM (i7 out of question, then) with upgrading to 256/512/1024 GB in mind for the next 2 or 3 years.

So here comes the idea:

• 2x Intel Xeon E5-2603 v3 (hexa-core, 1.6 GHz, 15 MB cache, 51 GB/s mem. bandwidth, quad channel each)
• 8x 16 GB DDR4 Kingston ValueRAM (2133 MHz, registered, ECC)
• Supermicro X10DRi motherboard

Our main concern revolves around these baseline Xeons. They're amazingly cheap (~200 EUR), have no hyperthreading (not needed), lower memory bus frequency (1600 MHz) and low processing power overall (6x 1.6 GHz = 9.6 GHz each), but they offer huge advantages on memory access (2x 51 Gb/s), cache (2x 15 MB) and maximum memory over a single more-powerfull Xeon or i7.

These components cost around 1700 EUR and the total for the workstation would be around 2200 EUR, with SDD, video card, PSU and case. That's cheaper than a single Xeon E5-1650 v3 with 4x 32 GB DDR4 and a Supermicro X10SRL-F (which I believe would be more powerful with 6x 3.5 GHz = 22.8 GHz, but with lower memory acess of 68 GB/s), and it's also able to handle a lot more upgrading in the years to come, both on processing power and memory capabilities. But are we missing something? Is there a bottleneck we're not seeing? If not, why aren't these baseline Xeons more common?

Thanks a lot in advance!

[flotus1]

First things first, the Xeon E5xxxxv3 is outdated by now as the v4 versions are already available.
The reason why these low-end Xeons are not widely used in Workstations should be quite obvious. They are slow. Like in REALLY slow. It is not only the 1.7GHz base frequency without turbo boost. The QPI speed is limited and of course they only support slow memory. Overall, you are wasting a lot of potential computing power. Keep in mind that with larger meshes, the copmuting power needed to run a job in the same amount of time increases.
My minimal recommendation for a professional CFD workstation would be the E5-2637v4 along with DDR4-2400 memory. If you really want to cheap out on the most important part of your new CFD workstation take the E5-2623v4.
Another reason why these low-end processors are not used when computing power is needed might be that buying used is the better option on a tight budget instead of buying low-end new hardware.

Edit: if still in doubt look at the cost for your CFD software licenses and compare them to the cost for your workstation. Spending a little more money to make better use of the expensive software is a no-brainer. By the way, how many licenses for how many cores do you have or are planning to buy? This basically determines what your ideal setup looks like.

[pagfetter]

Originally Posted by flotus1

First things first, the Xeon E5xxxxv3 is outdated by now as the v4 versions are already available.
The reason why these low-end Xeons are not widely used in Workstations should be quite obvious. They are slow. Like in REALLY slow. It is not only the 1.7GHz base frequency without turbo boost. The QPI speed is limited and of course they only support slow memory. Overall, you are wasting a lot of potential computing power. Keep in mind that with larger meshes, the copmuting power needed to run a job in the same amount of time increases.
My minimal recommendation for a professional CFD workstation would be the E5-2637v4 along with DDR4-2400 memory. If you really want to cheap out on the most important part of your new CFD workstation take the E5-2623v4.
Another reason why these low-end processors are not used when computing power is needed might be that buying used is the better option on a tight budget instead of buying low-end new hardware.

Edit: if still in doubt look at the cost for your CFD software licenses and compare them to the cost for your workstation. Spending a little more money to make better use of the expensive software is a no-brainer. By the way, how many licenses for how many cores do you have or are planning to buy? This basically determines what your ideal setup looks like.

First of all, thanks for the feedback, Alex!

So, sorry, my mistake, we're actually using OpenFOAM, I'll edit the post. Used to work with Fluent at my university's lab, little brain lag.

Considering normal licence costs per seat/cores, yes, faster CPUs will make sense. Very good reason not to use low-end processors. But for us it's really not the case, we just need a cheap rig to "leave it crunching big meshes". No big deal if it takes 20 or 30 hours instead of 5, in the end we won't launch more than one or two runs per week. For the rest of the time it will be used as just another CAD station.

As for the specific models: you're right, the new v4s cost almost the same. The E5-2637v4 would be way too expensive for us, although it would make sense with the licensing issue. E5-2623v4 may be possible, but I'm not sure yet if the higher price is worth the performance gains (~30% faster?).

On the other side, I gave a look on used hardware and found some pretty cheap pairs of E5-2620v3, even cheaper than new E5-2603v4s, which I think would fit like a glove in our tight budget and specific goals.

[RobertB]

Why not a 2620 which on SpecRate is 1.6x faster and $30 cheaper? If you don't need single threaded speed it would seem the best super cheap option for CFD. Make sure to get a V4 and DDR42400.