Per Core Performances of High End CPUs?

[pman27]

Hi Folks,

Forum newbie, my apologies if this is a topic revisited often, but your guidance would be appeciated.

I run a complex simulation program that first must do single core/single thread pre-processing (hyperthreading disabled), then can perform parallel processing, though not necessarily fully parallel and fully efficient. In large simulations, the single threaded processing can take several hours until the parallel processing takes over, which can then take several days.

Given this, we are looking for the highest performance CPU. The number 2 ranked CPU per this site's High End CPU Benchmarks is available to us - the Intel Xeon E5-2699 v4 @ 2.20GHz. It is 22 cores (though with 2 logical cores per physical core, I assume that is 11 physical cores with hyperthreading disabled). My question is whether this is the best CPU for us, where it achieves high benchmark scores due to parallel processing strength but the single physical core performance may not be ask strong as other machines with fewer physical cores?

In general, looking for a recommendation for the best performer for my single core pre-processing then multi-core parallel processing scenario.

Thanks in advance!

[David (PassMark)]

though not necessarily fully parallel

Do you have a budget? Or are just after the best of what is available?

Once in parallel mode, how many cores does it / can it use? i.e. would there be any benefit in getting a dual or quad CPU setup (40+ cores)?

Where is the bottleneck? Does it use the disk or networking at all? Does it make heavy use of RAM? No much point adding more cores if you don't have the RAM bandwidth?

Is the application NUMA aware, to support multiple RAM banks?

[pman27]

Hi David - thanks for the fast response. Looking for the best of what is available, cost not a major factor, though looking to keep the total system below $10K unless justified to go higher. For the moment we're talking single CPU though dual CPU could be considered. Not really assuming a bottleneck - the application uses both RAM and disk, so we're looking at getting 128 GB or 256 GB and as big an SDD as we can get.

We have a quote for an HP z640 single CPU machine with the Intel Xeon E5-2699 v4 @ 2.20GHz CPU, 128 GB of RAM, looking into larger SDD. Just not sure if that CPU will meet the bill performance wise, given the description of the application.

Not sure about NUMA aware. The number of cores used by the application is configurable, my application software guys say to take 40% of available memory and divide by 3GB to determine number of cores to allow. 64GB would mean 8 cores allowed for parallel processing, 128 GB would mean 16 to 17 cores.

Thanks in advance for your input.

[David (PassMark)]

If you are sure disk and RAM aren't a bottle neck and software supports NUMA have a look at the multi-cpu benchmark chart. Dual E5-2699 v4 is currently the top of the chart. Single threaded performance isn't great however. If what you are saying is true however about the RAM then you might not be able to get enough RAM into the box.

Motherboard like the Supermicro X10DAi might work. It supports 2TB of RAM.

For the disk get a M2 drive. Maybe a Samsung 960 Pro 512GB. Or if the motherboard doesn't support M2, then use a PCI-E based hard drive.

[pman27]

Hi David - what is your recommendation for a single CPU with a the best per core performance but a reasonably high number of physical cores (e.g. 8 to 12)?

[David (PassMark)]

If money is no problem then then Intel Core i9-7900X @ 3.30GHz

[pman27]

Thanks David - how would you expect it to compare to the Intel Xeon E5-2699 v4 @ 2.20GHz that we can get for my described application? Are my concerns about the E5 valid, with the i9 offering less exposure?

If the i9-7900X is not available for us through our supplier then do you have any reasonably similar suggestions? I've been consulting your http://www.cpubenchmark.net/high_end_cpus.html chart, but remain unclear on what might provide the best balanace of per core vs overall multi-core performance.

[i7Baby]

The amd Threadripper 1950x would be high on the list.