[MrAndersDk]

Hello

Currently I'm running OpenFOAM on a dell laptop workstation, with an i7-7820HQ 2.90 GHz cpu. At my firm we are looking into upgrading to a faster setup to run OpenFOAM on, but I find it difficult to evaluate benchmark for OpenFOAM use.

We are looking into two solutions at the moment. A budget version and a full version. For the budget version I'm thinking of buying the following setup used on ebay:

SUPERMICRO H8QGI-F
4x AMD 2.8GHz 16-Core Opteron 6386 SE
64 GB ram (8 x 8 GB) RDIMM 1600MHz
Quadro P2000
SAMSUNG EVO 860 1 TB
SEAGATE BARRACUDA 4TB

my estimate is that I can build such a system for rougly 3000 - 4000 USD here in Denmark.

The full version is ordering something like this

https://www.titancomputers.com/Titan...ion-p/w375.htm

upgraded with 2 x EPYC 7551 2.0 GHz 32 core CPUs and 64 GB ram. This setup will cost rougly 15000 USD here in denmark.

My problem is evaluating how the current setup will compare to these two.

The i7-7820HQ score rougly 10000 on cpubenchmark, the opteron also scores rougly 10000, and the EPYC scores rougly 18000. Ok since there are 4 opteron CPUs and 2 Epyc CPUs one could think that the opeteron system is 4 times better than the current setup, and the Epyc is only 3 times better. But I guess I can't use benchmarks that way to evaluate how they will perform with OpenFOAM.

So maybe someone in here can help me evaluate the system and compare the three systems?

And alternatively suggest other systems in this pricerange, if going for Xeon for example would give better performance for the money, even though as far as I can see will give a much lower core count for this price range.

[JBeilke]

Did you study this thread at all?


OpenFOAM benchmarks on various hardware

[MrAndersDk]

Yeah I tried to study that thread. But still find it difficult to evaluate the systems.

If we take the i7 8600 K it has 6 cores and a turbo freq of 4.7 GHz, multiplied this gives = 28.2.

The 2x Epyc 7501 has 64 cores and turbo of 3 GHz = 192.

A naive approach gives that the double Epyc should be 6 - 7 times better than the i7, in ok agreement with https://openfoamwiki.net/index.php/Benchmarks.

However, the same calculation for the 1950x does not work. It does not scale as well with the core numbers. Don't know if it has to do with the memory controllers they mentions.

If I anyhow assume I can do calculations like this, then the opteron calculation would be:

3.5 GHz * 64 = 224

making it surpass the double epyc. If this is anyway correct I guess it is a no brainer which system I should choose, however I don't think it is that simple.

I couldn't fin anything on opterons or quad cpus in the post, so don't know if it really scales that well.

[flotus1]

Even for compute-bound algorithms it is not that simple. Aggregate CPU frequency barely works when comparing CPUs of the same generation and architecture. It completely falls apart when comparing CPUs that are several generations apart.
A better estimate for CFD would be comparing memory bandwidth. But here again, the old Opteron CPUs will fall short. We already had some people here who just could not get them to work anywhere near the expected performance. 3000$ for such a system would be a total waste of money.

[MrAndersDk]

Ok. But the epyc system is a good value for money? Though it is clearly more expensive. Our should we consider another system?

Any systems in the 3000 usd range worth considering, were I will see an real increase in speed?

[gkarlsen]

Originally Posted by MrAndersDk

Ok. But the epyc system is a good value for money? Though it is clearly more expensive. Our should we consider another system?

Any systems in the 3000 usd range worth considering, were I will see an real increase in speed?

Why don't you just run the benchmark on your current system and then compare it to the results from others? https://openfoamwiki.net/index.php/Benchmarks

Since your current system appears to be dual channel, there are significant gains to be had. A 1920X build with 4 populated memory channels is well within your 3000 USD budget, and I would guess you will find it to be ~3 times faster than your current setup (dependent on cell count and solver). The 7820X is also good value if you prefer Intel. An EPYC build will give you 8 memory channels and a much higher speed, but it does come with a price premium.

[flotus1]

Not quite sure how that system could be 15000$. Anyway, buying from Titan computers you need a different configuration. They offer no configuration with 64GB of RAM and 16 DIMMs. So you need at least 128GB of RAM. The 32-core CPUs are pretty useless. Get 2x16 cores or 2x24 cores at most. Also, there are other vendors: https://www.deltacomputer.com/d20z-uln-zn.html
In the 2000$ range I would recommend a system with 2 Xeon E5 v2 CPUs. For example Xeon E5-2690 v2 or E5-2667 v2.

[MrAndersDk]

Thanks for the link to a Europe based vendor.

Think the dual epyc 7351 can be bought there for about 6000 usd. This is much more reasonable.

Now I just have to convince my boss the money is well spend. But since our current laptop are 4500 usd systems, hopefully it will not be hard

[Simbelmynë]

If you are running steady-state calculations similar to the one in the benchmark thread then I suggest you evaluate those results. It is easy to sort the list in a price performance chart (where you insert your local price).


The 7301 is very close to the 7351 if you consider similar memory speeds (the chart has not been updated). Hence there is little point in spending money on a more expensive EPYC setup. 7301 is the sweet-spot for CFD.


The best value is probably a cluster of E5-2690v2 such as the one by @spaceprop if you can find an 1866 MHz memory setup and a reasonable infiniband.


If you run mostly light- and/or single-threaded cases then the 8700k is best value.