CFD Online Discussion Forums

CFD Online Discussion Forums (http://www.cfd-online.com/Forums/)
-   Hardware (http://www.cfd-online.com/Forums/hardware/)
-   -   Opteron 63xx vs Xeon Memory Bandwith (http://www.cfd-online.com/Forums/hardware/122356-opteron-63xx-vs-xeon-memory-bandwith.html)

OICW August 17, 2013 19:43

Opteron 63xx vs Xeon Memory Bandwith
 
Hi guys,

Been reading the forum and learned lots of good info. I intend on building a system for our Fluids dept for Ansys and potentially OpenFoam to save on licensing cost.

So what I've learned is that memory bandwidth is critical for CFD work. The best bang for the buck is probably a cluster of 3930K PCs, but I'd like a 4 socket build to avoid the hassle any cost of Infiniband or 10Gb ethernet.

So my question: Why is everyone saying Xeon is much better than Opteron? Are they not both quad channel DDR3? Therefore 16 lanes total per 4p board?

I've looked all over the internet and can't find exact numbers for Opteron 63xx vs E5-4xxx memory bandwidth. What I do know is the Xeon would be at least double the cost.

Furthermore, seems like Xeons can only support DDR3-1600, while the Opteron is 1866. How can the Opteron be slower? Is something else bottlenecking it?

Thanks,

Roy

CapSizer August 19, 2013 08:24

There are several answers to your question, but if you look at the published specFPRate numbers for the CFD-related benchmarks for the various Xeons and Opterons, it is not clear at all that the Xeon necessarily has a performance advantage. In a given CFD application, this may or may not be true, dependent on a number of different factors, such as compilers optimizations. What you can reasonably conclude is that the extra cores of a 16 core Opteron vs. an 8-core Xeon do not translate into a big advantage for the Opteron. Either the performance is memory bandwidth constrained, or there is contention for the shared floating point units.

OICW August 19, 2013 21:27

Thanks. I guess I'm not looking for a detailed breakdown of Xeon vs Opteron performance, but rather, to answer this question:

Given: CFD is highly memory limited, and that both a 4p Xeon and 4p Opteron system have 16 lanes of DDR3 total.

Question: what's the cheapest way of flooding all 16 lanes of memory?

A 4p Xeon E5-4610 (2.4ghz hex) is $1400, an entry 16core Opteron is $700.

If both CPU can flood 4 lanes of DDR3 @ 1600 mhz, why would I want to double my system cost and get the Xeon?

I realize that Bulldozer/Piledriver cores are inferior to Sandy Bridge, but the ratio can't be that bad! Besides, if CPU is not the bottleneck in CFD, does it really matter?

CapSizer August 20, 2013 01:34

I don't think there is anything wrong with your logic, it is just always difficult to find real world comparative test results. FWIW, I've compared the performance of the basic cores involved (Intel: quad core i5, AMD:8-core 8150) on identical memory systems (dual channel DDR1600) and the results were close to identical. I * think * it is reasonable to scale this up ... an 8-core Xeon on 4 memory channels should be very similar to 2X the core i5, likewise 16-core Opteron should be similar to 2X8150, etc, then extend that argument to a quad-socket system ...

OICW August 21, 2013 14:34

Hmm, looks like I'll have to be the first one to find out...

Thanks for the help.

CapSizer August 21, 2013 15:35

Quote:

Originally Posted by OICW (Post 447254)
Hmm, looks like I'll have to be the first one to find out...

Good luck, and please share with us what you learn!


All times are GMT -4. The time now is 02:34.