[gmx-users] 64-bit vs. 32-bit

[Pim Schravendijk]

Hi Nguyen,

You should check the values on the benchmarking page, they've recently 
updated it, as well as the recommended cpus:

http://www.gromacs.org/benchmarks/single.php

My own experience:

On the 2.2 Ghz 64 bit AMD I can use here, specs see below, I have a
d.villin benchmark score of 9960 ps/day with a 32-bit binary, if I compare
this with the normal 1.6 AMD (6096 ps/day) I get an increase of only
3/2th.

model name      : AMD Athlon(tm) 64 Processor 3200+
cpu MHz         : 2202.917
cache size      : 512 KB

model name      : AMD Athlon(TM) MP 2000+
cpu MHz         : 1667.079
cache size      : 256 KB

On compiling: Compiling in 64 bit mode will make the program even slower
since it can't use the assembly loops for hardware graphic acceleration 
yet. Actually what I've done is just use the binary I've compiled on
the 1.6 ghz 32 bit AMD, it works fine, don't know if you can do this trick
with mpi as well, maybe if it uses static libraries on a fileserver that
can be accessed from both the 32 bit and the 64 bit machines (?).  
Otherwise you should install 32 bit libaries on the 64 bit machine and
mess around a bit to get it working.

Now 10 ns/day for d.villin is quite good, it's comparable with a 2.8 Xeon,
for example, and almost as fast as a g5 2.0 ghz mac with altivec loops
(this is also a 64 bit machine isn't it?). But what is happening at the
opteron with the fastest binary (32 bit) is that the fancy 64 "bitness"  
isn't really used. For example, the Espresso program for polymer
simulations that they use at the polymer theory group here (mpip in
mainz), which doesn't have assembly optimizations, gains a 5 time speed
increase by going from the 1.6 32 bit to the 2.2 64 bit in 64 bit mode. I
guess this doesn't mean that gromacs is slow on the opteron, but more that
it is really fast on a 32 bit machine :)

Erik once mentioned that even with 64 bit SSE assembly loops, gromacs
doesn't really get much faster:

http://www.gromacs.org/pipermail/gmx-users/2003-October/007412.html

Is it really impossible to have the same gain when going to 64 bit as you
see for non-assembly-optimized code? I am just curious on what is causing
this "barrier".

Greetings, Pim

--
Pim Schravendijk - PhD Student
Max Planck Institute for Polymer Research
http://www.mpip-mainz.mpg.de/~schraven/

> ------------------------------
> 
> Message: 3
> Date: Thu, 2 Sep 2004 19:13:35 +0200 (MEST)
> From: Nguyen Hoang Phuong <phuong at theochem.uni-frankfurt.de>
> Subject: [gmx-users] 64-bit vs. 32-bit
> To: Discussion list for GROMACS users <gmx-users at gromacs.org>
> Message-ID:
> 	<Pine.LNX.4.44.0409021900530.1856-100000 at marge.theochem.uni-frankfurt.de>
> 	
> Content-Type: TEXT/PLAIN; charset=US-ASCII
> 
> 
> Dear All,
> 
> I would like to ask if the new version 3.2.1 of GROMACS can run on 64 bit
> machine Opteron 244 or Opteron 1.6GHz ? If so, in general their
> performance is much better than the 32 bit machine AMD 2800+?
> 
> Thanks,
> 
> Phuong