Hi Foamers ! I ran two benc
Hi Foamers !
I ran two benchmarks on recent AMD and Intel CPUs and got some interesting results I wanted to share.
Case 1 (small, 351000 cells, 1000 Iterations):
Case 2 (medium, 830000 cells, 400 Iterations):
Both cases were decomposed with the hierachical algorithm. Depending on the decomposition sequence (xyz, yxz, etc.) the calculation times differed only by a few percent and the differences were assumed to be negligible.
The following histograms show the absolute calculation times (smaller means better) for single process, 2x, 4x and 8x parallelized cases. Due to significant overhead and a generally small cell number the small test case was only run 2x parallel. The white markers indicate ideal times for parallel simulation runs.
All machines are equipped with 667 MHz DDR2 RAM. 8x parallel runs were performed on dual quad-core boards.
It's pretty interesting to see that rather cheap "Desktop" CPUs like the AMD Phenom 9500 can compete with high end server CPUs like the Intel Xeon E Series (that costs approx. 5 times more than the AMD Phenom) when running a single process.
The pricier Xeons parallelize much better than the low cost AMDs but still anything beyond two processes seams to be a waste of ressources.
Hi Andras, very nice perfor
very nice performance comparison. However, do you have also some experience with Intel's Core 2 Extreme Processors (similar to Xeon Processor) on desktop machines?
Also, it is fascinating why the nominally much slower AMD Phenom numbers out the Xeon 3 GHz machine. Do you have any explanation for this?
Hi Stefan, The Xeon Machine
The Xeon Machines seem to have problems with the memory bandwidth. As you can see the Phenom's Hypertransport also gets 'saturated' when running more then two cores. There is no more speedup. I have also benchmarked an Intel Q6600 (QuadCore) which behaves better in parallel.
Here are the times (in seconds) for the Q6600 running the 'Medium' Case (830k Cells, turbFoam):
RAM: 667 MHz
1x ... 9859s (f=1.00)
2x ... 6398s (f=0.65)
4x ... 5624s (f=0.57)
As you can see from the figures the Intel Q6600 utilises multiple cores better than the AMD Phenom. For 800MHz RAM you get these values for the Q6600:
RAM: 800 MHz
1x ... 9503s (f=1.00)
2x ... 5958s (f=0.63)
4x ... 5031s (f=0.53)
To cut it short: The AMD Phenom is 10 to 15% faster than the Intel Q6600 for single core tasks but parallelises pretty badly.
I suppose the cheap AMD Phenom beats the high priced Xeons in floating point performance for certain operations... but this is just a guess...
A well designed mainboard will give you roughly 10 to 15% in performance boost. The same goes for faster RAM which is incredibly cheap at the moment.
|All times are GMT -4. The time now is 21:42.|