[ksmithgall]

Hello everyone,

I’m looking to build an entry-level workstation/server for some longer transient, multi-phase Openfoam simulations with a maximum mesh size of 10-15 million cells. I’m torn between an older dual socket xeon configuration or a newer single socket i7. Ideal budget is around $2000 including case, power supply, cooling and hard drives, but somewhat flexible.

Configuration 1:
CPU: 2 x Intel Xeon E5-2670 or E5-2680 or E5-2690
MOBO: ASRock - EP2C602-4L/D16 SSI EEB Dual-CPU LGA2011 or Intel S2600CP2J or Supermicro X9DR3-F
RAM: 8 x 8GB DDR3-1600 ECC

This configuration would have a total of 16 cores with 8 memory channels, but with a RAM speed of 1600MHz.

Configuration 2:
CPU: 1 x Intel i7-7820X
MoBo: ASRock X299 Taichi
RAM: 8 x 8GB DDR4-2666

This configuration would have 8 cores with 4 memory channels, but with a RAM speed of 2666MHz.

From reading this forum for a while I know that memory bandwidth often becomes the limiting factor for parallel performance. Do you think the first configuration is handicapped by the slower DDR3-1600 RAM compared to the second? Which do you think would be better for longer parallel simulations? Or, is there a better option I didn’t consider?

Thanks,
Kurt

[Simbelmynë]

I think both configurations are close which points to going with the 7820x system.


However, I would definately wait a couple of days (Aug 13) to look at prices of Threadripper 2.

[flotus1]

I have two remarks concerning the systems in question:
For the older dual-socket 2011 platform, I would use v2 CPUs. They don't break the bank, consume considerably less power and support DDR3-1866 memory. You don't have to buy expensive memory with that spec though. The motherboard you chose (Edit: ASRock, not the other ones) allows you to increase memory frequency up to the maximum supported frequency of the CPU. I could run both DDR3L-1333 and DDR3-1600 at 1866MT/s with this board. At least with memory from reputable brands like Samsung and Micron.

For the I7-7820x, use 4x16GB instead of 8x8GB. You can be certain to get dual-rank memory and it is easier to reach higher memory frequencies with less DIMM slots populated. Speaking of which: I highly recommend faster memory for this configuration, DDR4-3200 at the very least.

Personally, I would get the dual-socket workstation. It beats a single Skylake-X CPU in most parallel CFD applications. And you can save some money, especially on RAM.

[ksmithgall]

Thanks for the comments Alex, that's a good suggestion to go to the E5-2690v2 to run higher speed RAM while also reducing power consumption. It's also good to know your experience running a faster memory speed with that motherboard.

After reading through the benchmark post again it gives the impression that the Xeon processors would give noticeably better performance compared to the i7-7820X for the parallel tasks I'm looking for.

1820X ~ 0.62 iter/sec
E5-2690 ~ 0.77 iter/sec
E5-2690 v2 ~ 1.01 iter/sec

[Simbelmynë]

I do not agree.



2 x 2695 v2 is ~0.75 iter/s not 1 iter/s.



Also, you do not have 2695 v2, but rather 2690.

[flotus1]

Of course we can start to cherry-pick the results. There is always some variation.
For example, I achieved 0.89 iterations/s on a dual E5-2687w. The v2 versions will be faster.
The_Sle, after a lot of tweaking, managed 0.72 iterations per second on an I7-7820x. OpenFOAM benchmarks on various hardware
And indeed spaceprop posted results over 1 iterations per second on dual E5-2690v2 OpenFOAM benchmarks on various hardware

Xeon E5-2690v2 and E5-2667v2 both go for around 250$ these days. The latter has slightly higher clock speeds and cache per core, but two cores less. Both are a good choice.

[ksmithgall]

Here's where I pulled both the E5-2690 and E5-2690 v2 numbers from:
OpenFOAM benchmarks on various hardware

[Simbelmynë]

You are right, those numbers differs from the compiled table by quite a bit.


I guess the discrepancy is connected to the 1866 MHz memory used.


Good value anyways.