[GTCo8]

Originally Posted by Jordi

what's wrong with AMD?

Nothing is wrong.

I see that AMD FX 8150 (8 cores, Bulldozer) in some activities is comparable in speed to Intel i7-2600 (4 cores, Sandy Bridge) and it is only little cheaper than Intel. But only in some. Core per core is worse than Intel. Maybe this will change a little in improved Windows 8 and new Linux kernel. I have read that FX 8100 will be optimal (price vs speed).

Another problem with AMD MOBOs is that they support less memory (usually 4 banks, Intel have 6 or even 12 banks). I mean desktop MOBOs, not Server MOBOs.

And one thing more: MOBOs for AMD are cheaper than for Intel.

[wyldckat]

Greetings to all!

Just to give a small update on this subject:

• scipy posted some more numbers a while back: http://www.cfd-online.com/Forums/har...tml#post339168
• I picked up on this report: http://www.phoronix.com/scan.php?pag...r1_ofast&num=1 - this clearly indicates that not using the right options for compiling for the AMD Bulldozer, will affect the runtimes everyone is getting!

So, when in doubt, contact your CFD software supplier and ask them if and which versions of your favourite solver/meshers have been tuned for the AMD Bulldozer or any other CPU architectures!

Best regards,
Bruno

[daveatstyacht]

Scipy,
Thank you for going and benchmarking the two systems, I had not considered AMD as a possible candidate for my next CPU. The fact that overclocking the intel did not help your performance says to me that a different part of your system is bottlenecking the performance (probably the bus or memory). Interestingly, I went into researching the differences that might give the FX 8150 an edge and an area that it beats the 2700k is L2 cache hands down (2048 KB per two cores vs 256 KB per core or in other words 8 times more!). Since cache misses can have a huge effect on performance, the larger L2 cache could help significantly, particularly if something like multi-grid is being utilized. Another thing to consider is the maximum supported memory speed 2700k: 1333 MHz vs the 1866 MHz of the FX 8150. Memory speed is an important consideration for unstructured meshes. I think Bruno brings up an excellent point considering the potential bias that can be introduced by the compiler.

Dave

Sources:
Comparison of the two:

http://www.knowbytes.com/home/articl...us-amd-fx-8150

And to confirm the numbers are all correct go to the product pages:

http://www.amd.com/us/products/deskt...omparison.aspx
http://ark.intel.com/products/61275/...che-3_5-GHz%29

[rmh26]

The L2 and L3 cache's are shared on Intel's processors (smart cache) so a four core i7 would be able to share 1Mb of L2 and 8Mb of L3 while a FX4100 would have two independent 2Mb L2 caches(one per FPU unit) and a shared 8Mb L3. The completely shared cache would seem to give Intel an edge for smaller threaded problems while AMD would win out when working on larger sets.


I would say a big plus on AMD's side is the inclusion of FMA which would greatly speed up many linear algebra operations. I don't think Intel will get FMA until Haswell.

[mlotek]

What about cost vs performance? When I saw the prices of Intel processors...

I acquired x4 AMD Interlagos 6274's (64 cores) for under 250.00 USD each (granted they were used). I am currently running unstructured meshes in openFoam > 16,000,000 cells and find convergence times to be very reasonable (even with the memory bandwidth bottleneck...). I am considering getting another 64 cores because the cost is so low...

https://kudlaengineering.wordpress.com/

I would not rule out AMD, especially if its a cost vs performance issue. Now a watt/performance issue may be a different story.

Best,

Tom

[wyldckat]

Greetings Tom,

You didn't mention how many sockets you're using per motherboard. If you have one motherboard per processor, along with 4 DDR3 memory modules per motherboard, then it's only natural that you have a very good performance! Each processor is able to use 4 channel DDR3, therefore having a good configuration.

But you should also try to test using only 8 cores per processor and compare the performance you're getting. There was a post somewhere that described how to set the core affinity in mpirun... it's this thread: http://www.cfd-online.com/Forums/har...arameters.html

Best regards,
Bruno

[CapSizer]

Looks like a handy system Thomas. What did you do for CPU cooling, and how are the noise levels?

[mlotek]

Hi wlydckat,

All 6274's are on one board, H8QGi-F. The hard drive is a 256 gb SSD (very fast but way too small). I read about the performance penalty for unstructured grids due to the memory bandwidth (though this board has quad memory channels), but due to the used prices I found, I went for it anyway. So far, I am very happy with this setup and left space for a second board and switch (maybe infiniband if I can find a decent used one for not too much money, I have heard horror stories of Ethernet performance from HPC engineers).

Speed wise, I have not ran any benchmarks - it's been running OpenFOAM almost nonstop since I bought it and hasn't skipped a beat.

Hi CapSizer,
I used (4) Noctua fans (which are very quiet). Unfortunately, the two fans on the server case are very loud and will need to be replaced.

I would like to run some benchmarks, but not just with 8 cores...

Best,

Tom