[Simbelmynë]

Hey,


So over at https://openbenchmarking.org/test/pts/openfoam we have some nice up-to-date information about current CFD hardware. I am curious what your thoughts are when it comes to the 30M vs 60M test cases, especially in the lower tiers (e.g. 5900X vs 12900k). 5900X performs reasonably well at 30M cells, but for 60M cells it is surpassed even by the 7700k. This really does not make sense to me. Does the extra cache give benefits for the 30M case but not for the 60M case, or is it something else?

[flotus1]

This site has come up from time to time. And each time I looked at it, I was just baffled by the results. Many of them just don't make any sense, as you found out yourself. Not sure what to make of it
Some of the oddities you describe could be explained by different memory configurations. I.e. simply not enough memory on one of the systems to run the 60M case in-core. But then again, what's the point of keeping such results.

[Simbelmynë]

Yes I agree. Strange though that there are many such reported results and that the spread is quite low.



Another thing that seems a bit suspicious is the number of CPUs. I have a hard time believing a single 7763 is as good as 2 x 7453. Perhaps it is manual user input that is behind such reports? The idea is good though and the overview is much better compared to our in-house solution here at cfd-online

[flotus1]

The Epyc 7453 is a special case. With only 64MB of L3 cache, it is the weakest CPU in the whole lineup. Possibly with only 2 CCDs enabled, and thus reduced effective memory bandwidth. A single 7763 has to be faster than 2 of these.

The main issue is probably that these results are all crowd-sourced. And I don't know how much effort goes into verifying the results uploaded to the database. What's the memory configuration? Bios settings? Operating system?
There are just too many variables in play. Benchmarking with consistent results isn't easy, even people who do it for a living get it wrong sometimes.

[Simbelmynë]

2 x 7F72 is only about 10% faster than the 7763. So although I hear what you say about the gimped 7453, I still think this is strange, unless perhaps Milan is really that good. Also, the standard deviation is really low for many of the benchmarks. I am not sure if that indicates that the same person has submitted several benchmarks for the same system. We have much larger variations in our benchmark thread here, with respect to the variables you mention.

[aparangement]

I've tried (half a year ago?) to run its benchmark by myself.

However even the "30M" case would use up the 256G RAMduring snappy mesh. I tried to setup a coarser mesh however my wall time is not comparable at all with the his report, even compare with the 30M setup (something like 20 minutes vs 30 seceonds, maybe the reported endtime is not the same as the clean case provieded).

Then I gave up digging...

I agree that the results are not in consistent with each other, althogh I don't have proof..
I highly suspect that some of the too fast cases are due to crashed simulaiton..

Originally Posted by Simbelmynë

Hey,


So over at https://openbenchmarking.org/test/pts/openfoam we have some nice up-to-date information about current CFD hardware. I am curious what your thoughts are when it comes to the 30M vs 60M test cases, especially in the lower tiers (e.g. 5900X vs 12900k). 5900X performs reasonably well at 30M cells, but for 60M cells it is surpassed even by the 7700k. This really does not make sense to me. Does the extra cache give benefits for the 30M case but not for the 60M case, or is it something else?

[Ford Prefect]

Yeah something is a bit fishy. If there are crashed cases I would have guessed that the standard deviation would be larger though. But I think that is a plausible reason for the strange results.

[linuxguy123]

Originally Posted by flotus1

The main issue is probably that these results are all crowd-sourced. And I don't know how much effort goes into verifying the results uploaded to the database. What's the memory configuration? Bios settings? Operating system?
There are just too many variables in play. Benchmarking with consistent results isn't easy, even people who do it for a living get it wrong sometimes.

So true. Want proof ? Go over to www.geekbench.com and compare the ratings people get on the same processor. It is not unusual to find a 2:1 rating difference between the fastest and the slowest.


Great discussion.