[Nautilus1954]

Hello,

I have one question related to processor selection for Fluent simulation.

Intel Phi processor often has more cores, but lower frequency than Intel® Xeon® Processor E7 processor.

For example, in order to have 64cores, one Intel Phi processor can be enough, but 4 Xeon® Processor E7 processors are needed. Most likely 4 Xeon® Processor E7 processors solution is more expensive. But can 4 Xeon® Processor E7 processors solution also promise higher Fluent simulation speed?

How to compare the Intel Phi processor frequency and Intel® Xeon® Processor E7 processor frequency?

Any advice is appreciated. Thanks in advance.

[flotus1]

Both Intel and Ansys appear to be working on this topic together. However, the fact that it is not advertised aggressively should be our first clue that Ansys with KNL is not yet competitive. See for example this blog post here: http://www.ansys-blog.com/boost-ansy...-technologies/
If even Ansys admits that they are not yet satisfied with the performance, the customer should definitely avoid the feature for now.
You should instead consider the first point on their list: Intels new Skylake-SP processors. They replaced the older Xeon E5 and E7 processors.
Or you can have a look at what AMD has to offer with their new "Epyc" processors.

Comparing KNL and a "regular" CPU is next to impossible. The differences in the architecture are quite large. See For example this presentation if you want to learn more: https://colfaxresearch.com/how-knl/

[Nautilus1954]

Thank you very much. This is a very valuable point.
The Intel processors based on Skylake-SP processors now are mainly Xeon Scalable Processors. I will look deep into those products.

[kyle]

Xeon Phi's potential for CFD is not the huge core count, it's the 16GB level 4 cache. If you use this cache in conjunction with the main system memory, you get about 500GB/s of memory bandwidth. That is over 4x as much as the newest Xeon Gold. Of course the catch is your simulation needs to be small enough to fit in the cache.

The huge core count is actually a bad thing for CFD since it creates a lot of inter-process communication overhead. If Intel would release a more traditional 8-16 core processor with this same 16GB cache, that would be the ultimate CFD processor.

Here is a benchmark that shows a single Xeon Phi as significantly faster than two last-gen Xeon E5's: https://www.ictgmbh.net/boost-openfo...ntel-xeon-phi/

[JBeilke]

It would be interesting to find out the best option for a 8 core configuration.

• one Epyc
• two Epycs
• Skylake-X
• Threadripper
• Xeon-Phi

[flotus1]

If we are talking about maximum performance in a single machine using off-the-shelf hardware while being limited by an 8-core license, my money would be on a quad-socket system with 4 Xeon Gold 6144. The frequency when running 2 cores per CPU should be 4GHz or higher and the L3 cache is relatively large.
Of course compared to other possible solutions you are spending several times more money for just a few percent more performance. On the bright side however, you are prepared if you ever upgrade your license to 32 cores

[JBeilke]

Originally Posted by flotus1

If we are talking about maximum performance in a single machine using off-the-shelf hardware while being limited by an 8-core license, my money would be on a quad-socket system with 4 Xeon Gold 6144. The frequency when running 2 cores per CPU should be 4GHz or higher and the L3 cache is relatively large.
Of course compared to other possible solutions you are spending several times more money for just a few percent more performance. On the bright side however, you are prepared if you ever upgrade your license to 32 cores

Ok, let's be a bit more realistic:

• You are the one who has to pay for the license (ccm+) and the machine.
• The machine will stay under YOUR desk. So noise and cooling (it should also work in summer) are also important.
• The tesla generator is still in developement, so we have to buy the electricity.

[kyle]

You've got to have an overall budget to answer that question. Incremental dollars will make it incrementally faster. You need to know what dollar total to stop at, and the optimal platform (Skylake/EPYC, dual/single socket, etc) will not be the same for every budget.

[flotus1]

This. A more "reasonable" solution would be a dual-socket workstation with 2 Xeon Gold 6128 (6 cores) or 6144 (8 cores). This way you can run a simulation on 8 cores and still use the workstation for other tasks like pre/postprocessing.
There is no point in going for a much cheaper option since the license cost per year is much higher than the price for the workstation. Same applies to electricity costs. And it is not that hard to get silent cooling for such a dual-socket workstation.

The other options:

• one Epyc: great value, but lacks performance per core compared to the Xeons
• two Epycs: same. Excellent choice if you do not have an expensive per-core licensing system
• Skylake-X: lacks memory performance with only 4 channels compared to 2x6 channels
• Threadripper: all of the above
• Xeon-Phi: lacks performance per core

[JBeilke]

Thanks Alex. There are rumours, that they want to change the licensing system for ccm+, but as long as it works this way, the 2x Xeon seems to be the best option.

I hope we get some benchmark data in the near future.

Cooling the machine is probably not the problem, but my office will become a little warm when running a 4 processor machine :-)