Intel E5-2698 16core CPU

AndyR · May 17, 2016, 11:36

Folks,
Does anyone have any experience with the Intel 16 core CPUs?
Specifically we are looking at the E5-2698 v3 16 core 2.3GHz
Two sockets per motherboard for a 32 core compute node.

Our experience with the 16 core AMDs was not a good one.
The fact that every pair of cores had to share a floating point unit and every 4 cores had to share a memory channel, slowed the machine down quite a bit.

I am concerned that we may run into a similar issue with the Intel

Any help is appreciated
Thanks
Andy

flotus1 · May 17, 2016, 12:30

The problem with shared execution units does not exist with the Xeon CPUs.
The bottleneck will be a different one: Memory bandwidth. The CPU only has 4 memory channels which can be a limiting factor for CFD simulations on 16 cores. Unless you know that your applications are not bandwidth limited the E5-2698 v3 (v4 is available) might not be the ideal choice
For a typical CFD compute node you should rather go for the "frequency optimized" versions of the Xeon Processors with less cores. Xeon E5-2687W v4 is the latest and greatest.
https://www.microway.com/knowledge-c...ep-processors/

RobertB · May 18, 2016, 09:10

Andy,

Go check out the SpecRates on the Spec.org site. For our last server they gave a good estimate of relative speed.

https://www.spec.org/cpu2006/results...2/#SPECfp_rate

We are looking at a smaller system and may go 12 core. The newest V4 cores have DDR4 2400 which helps bandwidth and they are 15-20% faster than the V3s (at least in this CPU number).

The CPU price varies quite a lot between quite similar CPUs so that is also a factor.

We found CPU locking was a significant factor in our present cluster for CCM+.

Robert

May 17, 2016, 11:36	Intel E5-2698 16core CPU	#1
AndyR Member Andy Robertson Join Date: Mar 2009 Location: Long Island NY Posts: 46 Rep Power: 17	Folks, Does anyone have any experience with the Intel 16 core CPUs? Specifically we are looking at the E5-2698 v3 16 core 2.3GHz Two sockets per motherboard for a 32 core compute node. Our experience with the 16 core AMDs was not a good one. The fact that every pair of cores had to share a floating point unit and every 4 cores had to share a memory channel, slowed the machine down quite a bit. I am concerned that we may run into a similar issue with the Intel Any help is appreciated Thanks Andy

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Superlinear speedup in OpenFOAM 13	msrinath80	OpenFOAM Running, Solving & CFD	18	March 3, 2015 05:36
Star cd es-ice solver error	ernarasimman	STAR-CD	2	September 12, 2014 00:01
[OpenFOAM] Color display problem to view OpenFOAM results.	Sargam05	ParaView	16	May 11, 2013 00:10
CFX11 + Fortran compiler ?	Mohan	CFX	20	March 30, 2011 18:56
OpenFOAM 13 Intel quadcore parallel results	msrinath80	OpenFOAM Running, Solving & CFD	13	February 5, 2008 05:26

May 17, 2016, 12:30		#2
flotus1 Super Moderator Alex Join Date: Jun 2012 Location: Germany Posts: 3,400 Rep Power: 47	The problem with shared execution units does not exist with the Xeon CPUs. The bottleneck will be a different one: Memory bandwidth. The CPU only has 4 memory channels which can be a limiting factor for CFD simulations on 16 cores. Unless you know that your applications are not bandwidth limited the E5-2698 v3 (v4 is available) might not be the ideal choice For a typical CFD compute node you should rather go for the "frequency optimized" versions of the Xeon Processors with less cores. Xeon E5-2687W v4 is the latest and greatest. https://www.microway.com/knowledge-c...ep-processors/

May 18, 2016, 09:10		#3
RobertB Senior Member Robert Join Date: Jun 2010 Posts: 117 Rep Power: 16	Andy, Go check out the SpecRates on the Spec.org site. For our last server they gave a good estimate of relative speed. https://www.spec.org/cpu2006/results...2/#SPECfp_rate We are looking at a smaller system and may go 12 core. The newest V4 cores have DDR4 2400 which helps bandwidth and they are 15-20% faster than the V3s (at least in this CPU number). The CPU price varies quite a lot between quite similar CPUs so that is also a factor. We found CPU locking was a significant factor in our present cluster for CCM+. Robert