AMBER Archive (2009)

Subject: RE: [AMBER] AMBER efficiency: Intel vs. AMD

From: Ross Walker (ross_at_rosswalker.co.uk)
Date: Mon Jan 19 2009 - 09:39:33 CST

Hi Swarup,

Ye Mei makes some good comments that indeed it is largely clock speed that
matters. Although getting higher speed memory and a faster FSB can help as
well although this typically adds a lot to the price for only marginal gain.

Here are some current performance numbers I have (more details are on
http://ambermd.org/amber10.bench1.html).

Throughput (ns/day) for various benchmarks. AMD Machine = TACC Ranger,
Barcelona 4x4core 2.3GHz. Intel Machines = 2x4core E5430 2.66GHz, 1333MHz
FSB, PC2-5300 FB ECC DIMMS & E5462, 2.80GHz, 1600MHz FSB, PC2-6400 FB ECC
DIMMS.

This is not a perfect comparison since the Ranger nodes are 16 way so the 8
processor runs are actually 2 mpi threads per core - expect slightly less
with dual core nodes since you saturate things quicker. It's also probably
not the best combination to compare but it was all I had access to. Anyway:

ns/day (8 processors)
Benchmark Ranger E5430 E5462
JAC NVE 5.01 6.09 6.70
FactorIX NVE 1.27 1.44 1.62

The clock speed ratio between the E5462 and E5430 chips is 1.053 whereas the
performance ratios are 1.107 JAC and 1.125 FactorIX so you can see that
memory bandwidth does help a bit - the Nehalem design should help a bit more
with this but don't expect any miracles. Between the E5430 and Ranger
machines the clock speed ratio is 1.157 whereas the performance ratio is
1.216 JAC and 1.134. So for JAC it likely benefits from the larger cache on
the Intel chip. The FactorIX benchmark is roughly the difference in clock
speeds. Remember though that since Ranger is 4x4core it's 8 way performance
is better than you would get out of a 2x4core machine.

Ultimately though the performance differences are not huge so I would
suggest going for the highest clock speed and memory speed you can which
means the Intel chips.

Then of course it simply comes down to a matter of price which of course
muddies the water significantly...

Good luck,
Ross

> -----Original Message-----
> From: amber-bounces_at_ambermd.org [mailto:amber-bounces_at_ambermd.org] On
> Behalf Of Swarup Gupta
> Sent: Sunday, January 18, 2009 5:44 PM
> To: amber_at_ambermd.org
> Subject: [AMBER] AMBER efficiency: Intel vs. AMD
>
> I wish to know if anyone has any suggessions whether intel or amd is
> preferred for a computing node with 8 cpus? My guess is that for intel
> there is a common shared memory whereas for
> AMD, each chip has its own memory controller - meaning that perhaps intel
> may have problems read/write during MD simulations with AMBER. Please let
> me know if anybody has any comments.regards,SGupta
>
>
>
> _______________________________________________
> AMBER mailing list
> AMBER_at_ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber

_______________________________________________
AMBER mailing list
AMBER_at_ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber