[gmx-users] Question about benchmarks

[Erik Lindahl]

Susan Chacko wrote:
> Hi folks,
> 
> Are the benchmarks on the www.gromacs.org site for single-precision
> builds of Gromacs, and would one expect a large difference between
> single-precision and double-precision times?
> 
> I just built Gromacs in double-precision mode, but am wondering
> whether I should go back and rebuild in single-p as well, as the
> benchmark times I'm seeing are quite different from those on the
> Gromacs website.
> 
> I'm running the DPPC benchmark on a Linux cluster, dual-processor Athlon
> 1.4GHz nodes, RedHat 7, MPICH.
> 

Hi Susan,

The benchmarks are done in single precision, since that is accurate
enough for almost everything except normal mode analysis.

For an athlon you will probably gain a factor of 2 in speed by using 
single precision; we have written assembly loops that use multimedia 
instructions, but Athlon & Pentium 3 CPUs only support this in single 
precision.

(On Pentium 4 systems we can do double precision multimedia 
instructions, but that is still 25-40% slower than single).

The same thing applies for Apple machines since we can use 
single-precision Altivec instructions, but for the rest (IBM, etc), the
difference is much smaller.

In general, I'd recommend to stick to single precision unless you are
really sure you need double.

Cheers,

Erik