CFD Online Discussion Forums

CFD Online Discussion Forums (http://www.cfd-online.com/Forums/)
-   Hardware (http://www.cfd-online.com/Forums/hardware/)
-   -   Anyone with Opteron 6272 / 6274 experience? (http://www.cfd-online.com/Forums/hardware/98569-anyone-opteron-6272-6274-experience.html)

CapSizer March 14, 2012 01:37

Anyone with Opteron 6272 / 6274 experience?
 
The SpecFPrate scores for a 4-way / 64 core Opteron 6272 or 6274 system look excellent (around 640, for those who look at this sort of thing). However, is there anybody with any real CFD experience of one of these monsters? It looks like a compact and affordable way of putting together a very powerful system, but I am concerned about these new cores with the shared fp processor. Is it performing in the real world? I don't want any "Intel better" comments, what I want to know is if the real-world performance is consistent with the SpecFPrate score.

kyle March 14, 2012 15:44

Quote:

Originally Posted by CapSizer (Post 349320)
I don't want any "Intel better" comments

Well, what if Intel is better? I have no idea if it is for your application, but that kind of attitude is unwise. I bought test nodes for each Intel and AMD systems, and for my application, Intel was far cheaper from a seconds-per-iteration standpoint.

The only way to know for sure which is best is to try it out with whatever code, mesh, physics and geometry will be typical of your runs. Ideal hardware will vary wildly with those four variables.

SpecFPrate is unlikely to be indicative of real world CFD anyway. Most traditional codes are bottlenecked by cache performance and memory bandwidth, not floating point performance.

CapSizer March 14, 2012 17:12

Quote:

Originally Posted by kyle (Post 349483)
Well, what if Intel is better? I have no idea if it is for your application, but that kind of attitude is unwise. I bought test nodes for each Intel and AMD systems, and for my application, Intel was far cheaper from a seconds-per-iteration standpoint.

SpecFPrate is unlikely to be indicative of real world CFD anyway. Most traditional codes are bottlenecked by cache performance and memory bandwidth, not floating point performance.

Well, if one can prove that Brand I is more cost effective than Brand A, you obviously buy Brand I. However, this kind of relevant information is insanely hard to come by, which is why I asked if anyone out there had any real information .... Most so-called benchmarks run by hardware review sites are totally meaningless, but SpecFPRate is a bit more sophisticated, if by no means representative of real-world CFD experience. I'm only in the market for a single system, so "buying a test node" is out of the question. The 6272/6274 family CPU's represent a major design philosophy change, and it would be very handy to know what it does in practice before rejecting or embracing it outright. So do you have any numbers?

scipy March 14, 2012 20:30

I can't pass judgment on the Opteron, but I did do a comparison between a i7-2600K and a FX-8150 desktop cpu. Since both the FX and the Opteron are of the same Bulldozer architecture, maybe some comparison can be drawn there.

All in all, the FX-8150 at clock for clock was ~13 % faster than the i7. Both are 8 core cpus (even though discussion can be had on the FPU/core and the i7 only has 4 physical cores and HT), maybe it's better to say I ran 8 Fluent threads on both and the AMD came out on top.

I'm quite sure that a i7-970 or something else with 6 cores would've ate the AMD because of the extra physical cores, but then again.. the Opteron does have double the cores of the FX-8150. However, the new E5 Xeon might be something to look into.. double the Memory bandwith of the i7, 8 physical cores for the first time, but the price is almost 4x the Opteron for a single Intel CPU.

If you have a local hardware vendor who would let you do an hour of testing at their place of buisness, the whole thing can be tested quite easily. After all, you would be leaving a rather large sum of money with them..

rmh26 March 14, 2012 20:51

http://www.anandtech.com/show/5553/t...for-servers/12

The new xeons do seem to have a per core/bulldozer module advantage but AMD makes up for it by being cheaper. How much are the mother boards for a quad socket G34. I only see one on newegg for $800. There are some reviews out there for the new opterons showing them not performing that well, slighty better or worse than the previous generation. I think the one area the do well in is HPC but it will be software specific.

My advice would be to check out an E5-2630. You could try talking to scipy, I think he was planning on buying a dual 6272 setup.

RobertB March 15, 2012 13:23

@scipy

Did you try the intel with only 4 threads and HT turned off?

Our experience with hyperthreading was bad, the OS did not understand which were 'real' cores and which virtual. On an 'underloaded' system (as many threads as real cores) it would schedule threads on the real and virtual cores of the same processor which obviously was not helpful. Turning off hyperthreading almost doubled performance in this case.

In your case all 'cores' are being assigned. However the OS may still be switching cores more than is desirable causing cache misses, etc. It would be interesting to hear about a 4 thread run with HT turned off to see if that changes the ranking.

scipy March 16, 2012 11:47

Quote:

Originally Posted by RobertB (Post 349679)
@scipy

Did you try the intel with only 4 threads and HT turned off?

Our experience with hyperthreading was bad, the OS did not understand which were 'real' cores and which virtual.

Yes, I've tried it. All the results are summarized in one of the posts in the i7 vs FX thread.


All times are GMT -4. The time now is 21:04.