[sisi]

I know that this is a frequently asked question. Which CPU is better for OpenFOAM: Intel i7-4770K or AMD FX-8350? If AMD may not be the best choice, would that be an acceptable solution if saving money is a concern? Or is the i7-4770K definitely worth the extra money (about 120 Euro)?

What is most important for CFD performance?

sisi

[evcelica]

If you are purchasing a new system you might want to consider the i7-4820K instead. It has 4 memory channels as opposed to only 2 on the CPUs you listed. The price may not be that much more than the 4770K system.

[natty_king]

OpenFOAM can only run multi-core on real cores, it cannot use hyperthreading.

Therefore if you are running in parallel you will only be able to run it on 4 cores on the i7-4770k but 8 cores on the 8350.

The i7 is stronger per core, but not 2x stronger. From personal experience the 8350 will run about 1.6-1.7 times faster than the i7 utilizing all cores on each.

Hope this helps!

[sisi]

Hi all, thanks for your replies! Yes, I am building up a completely new system. And I have never build up a system for a special software like OpenFOAM. So that was helpfull!

Originally Posted by evcelica

If you are purchasing a new system you might want to consider the i7-4820K instead. It has 4 memory channels as opposed to only 2 on the CPUs you listed. The price may not be that much more than the 4770K system.

If I have understood correctly Intel's i7-4820K does not make sense in combination with OpenFOAM, either? Am I right? LGA2011 Sockets are so expensive

sisi

[kyle]

i7-4820K is almost certainly the fastest chip being discussed if we are talking about CFD with an unstructured grid. Number of memory channels and memory latency is much more important than clock speed or number of cores.

People that suggest AMD chips are just looking at random general benchmarks on the internet. They have not run CFD-specific benchmarks.

[wyldckat]

Originally Posted by kyle

People that suggest AMD chips are just looking at random general benchmarks on the internet. They have not run CFD-specific benchmarks.

Check post #11 on this thread: http://www.cfd-online.com/Forums/har...tml#post336993 - it's a now somewhat old post where scipy details how going with an i7-2600K was the wrong choice back then, when compared to an AMD FX-8150, even though most benchmarks stated that the i7 was better. The detail back then was that both were using dual-channel memory controllers, therefore more real cores was better than using hyperthreading.

Nonetheless, as Kyle stated, the i7-4820K does have the upper hand here, since it has a quad-channel memory controller, which can boost considerably the performance.

By the way, there was post not long ago about how the memory speed affected the performance... ah, this one is a good reference: http://www.cfd-online.com/Forums/har...tml#post483489 post #5

[sisi]

Sorry for my late reply. Thanks for all your answers and the links! So the i7-4820k is better performing then the i7-4770k and the AMD FX-3850. Personally, I think that I don't need an expensive LGA 2011 Socket to start exploring the OpenFOAM environment.

Originally Posted by natty_king

OpenFOAM can only run multi-core on real cores, it cannot use hyperthreading.

Therefore if you are running in parallel you will only be able to run it on 4 cores on the i7-4770k but 8 cores on the 8350.

The i7 is stronger per core, but not 2x stronger. From personal experience the 8350 will run about 1.6-1.7 times faster than the i7 utilizing all cores on each.

Hope this helps!

Alright. That convinced me to go with the FX-3850 then.

sisi

[mk23]

i have compared the FX8130 on my desktop and I7 4700QM on my laptop with Ansys CFX Benchmark.def. 8130@4.5GHz OC is about 50% faster than 4700QM@3.5Ghz.
FX8130 22 sec
I7 4700 33 sec
@3.5Ghz they are both almost the same fast. So i would say for CFD core number is more important than performance per core.

[sharonyue]

Originally Posted by kyle

i7-4820K is almost certainly the fastest chip being discussed if we are talking about CFD with an unstructured grid. Number of memory channels and memory latency is much more important than clock speed or number of cores.

People that suggest AMD chips are just looking at random general benchmarks on the internet. They have not run CFD-specific benchmarks.

Im curious on this,

4820K VS 5820K

http://ark.intel.com/products/77781/...up-to-3_90-GHz

http://ark.intel.com/products/82932/...up-to-3_60-GHz

this price is almost the same. 5820k has 6 cores, cpu clock is 3.3. 4820k has 4 core, cpu clock is 3.7.

And the max memory bandwidth is 68GB/s for 5820 and 59 GB/s for 4820k.

Which one do u recommend?

[wyldckat]

Greetings to all!

Originally Posted by sharonyue

this price is almost the same. 5820k has 6 cores, cpu clock is 3.3. 4820k has 4 core, cpu clock is 3.7.

And the max memory bandwidth is 68GB/s for 5820 and 59 GB/s for 4820k.

I've been trying to come up with an heuristic to do an estimate of the performance of processors, based on their specs, for CFD purposes. So here's another attempt at analysing things.

Essentially you're comparing two CPUs of identical Lithography (22nm), where the 5820k came out during the Tock phase: http://en.wikipedia.org/wiki/Intel_Tick-Tock - which means that the 5820k is meant to be better by definition, because some fine tuning went into the development of a tock phase.

Now let's compare theoretical performance, based on the stats:

• i7 4820k:
  • 4 cores
  • 3.7 to 3.9 GHz - let's assume that with HT turned off, it can do 3.9 GHz, even if a little OC has to be done.
  • 59.7 GB/s maximum bandwidth, in other words, using 4 RAM modules, DDR3 1866MHz
  • Theoretical total CPU GHz: 4 * 3.9 = 15.6 GHz
  • Total memory bandwidth available per core: 59.7 / 4 = 14.925 GB/s
• i7 5820k:
  • 6 cores
  • 3.3 to 3.6 GHz - let's assume that with HT turned off, it can do 3.6 GHz, even if a little OC has to be done.
  • 68 GB/s max bandwidth, with 4 RAM modules, DDR4 2133 MHz
  • Theoretical total CPU GHz: 6 * 3.6 = 21.6 GHz
  • Total memory bandwidth available per core: 68 / 6 = 11.33(3) GB/s

Therefore:

• The total computing power in GHz gives a ratio of 1.38 in favour of the i7 5820k (5820k/4820k).
• But when taking into account the ratio of memory performance per core, it rates roughly 0.76 against the i7 5820k (again, 5820k/4820k).
• In theory, this would indicate that the i7 5820k is at best 1.38 * 0.75 = 1.035 faster than the i7 4820k. And no, this is not an accurate assessment, because this doesn't take into account several other details.

If you compare a few conventional benchmarks between these two processors, a few rate them as the i7 5820k being 1.31 times faster than the i7 4820k; but those benchmarks usually don't take into account the operations more common to CFD applications, where memory access is very important.


A more accurate possible heuristic would be to think something like this:

1. Thinking of maximum limits, we need to think about the amount of RAM available, since that will be used for the calculations.
2. Both are rated at a maximum of 64GB. This means that:
   1. i7 4820k => 64 / 4 = 16 GB per core.
   2. i7 5820k => 64 / 6 = 10.66(6) GB per core.
3. This means that each core has to handle data at the following rates:
   1. i7 4820k
      1. 16 GB / 14.925 GB/s = 1.072 s to go through the whole associated memory for each core.
      2. Processing 3.9 GHz => 3.9 * 1.072 = 4.1808 G-cycles
   2. i7 5820k
      1. 10.66 GB / 11.33 GB/s = 0.94 s
      2. Processing 3.6 GHz => 3.6 * 0.94 = 3.384 G-cycles
4. This doesn't feel very correct, but the theoretical scale up would be 4.1808 / 3.384 = 1.235

I still need to do some more math on this topic. If you look for:

OpenFOAM xeon benchmark

there are already a few that can be used to do some theoretical heuristic calculations/validations on this topic.


Best regards,
Bruno

[sriny1512]

Dear Friends,

I have done a Computational Fluid Dynamics problem solving benchmark on 3 different computers. The problem solved while benchmarking is a turbomachinery problem. I hope this data is useful to people looking for CFD bechnmark in particular. I am sorry, I do not have HPC access to use all cores. Below are the details tabulated. Thank you.