AMD Multicore + CUDA Benchmarks, are them ok ?

From: David McGiven (davidmcgivenn_at_gmail.com)
Date: Thu May 05 2011 - 04:08:30 CDT

Dear NAMD Users,

We have recently bought a multicore-AMD (4 x 12-core Opteron 6176 2,3
Ghz) server with a Tesla S2050 (4 GPU) and I would like to know if
this benchmarks are ok :

The test is the standard apoa1.namd (downloaded from : http://www.ks.uiuc.edu/Research/namd/utilities/apoa1/)
  plus I've added this three lines (as NAMD required them) :

outputEnergies 1000
xstFreq 1000
dcdFreq 1000

I got the Benchmark times from NAMD output i.e :
Info: Benchmark time: 12 CPUs 0.0301262 s/step 0.348683 days/ns
61.6245 MB memory

I get the following results for CPU :

1 CPU cores : 20,629 days/ns
2 CPU cores : 10,789 days/ns
24 CPU cores : 0,942 days/ns
48 CPU cores : 0,494 days/ns

I get the following results for GPU :

1 CPU core + 1 GPU : 2,231 days/ns
2 CPU core + 1 GPU : 1,250 days/ns
6 CPU core + 1 GPU : 0,630 days/ns (best result, increasing cpu cores
decreases performance)

1 CPU core + 2 GPU : 2,27 days/ns
2 CPU core + 2 GPU : 1,28 days/ns
10 CPU core + 2 GPU : 0,47 days/ns (best result, increasing cpu cores
decreases performance)

1 CPU core + 3 GPU : 2,27 days/ns
2 CPU core + 3 GPU : 1,264 days/ns
12 CPU core + 3 GPU : 0,38 days/ns (best result, increasing cpu cores
decreases performance)

1 CPU core + 4 GPU : 2,18 days/ns
2 CPU core + 4 GPU : 1,292 days/ns
12 CPU core + 4 GPU : 0,34 days/ns (best result, increasing cpu cores
decreases performance)

What do you think ? Are these the expected results for a Tesla ? I am
a little bit disappointed by the performance because AMD machine
costed the same as the Tesla, for getting almost the same
performance ...

Well, thanks in advance.

Regards,
David

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 05:23:53 CST