[gmx-users] Benchmarks, two questions

[Erik Lindahl]

> - How come the 32 bit gromacs binaries score relatively so well on the 
> 64
> bit machines. Is there some special way in which the processor treats 
> the
> 32 bit operations?

SSE assembly loops.

I've actually translated them to 64-bit SSE, but it doesn't make things 
any faster per se, so that's why I haven't bothered to backport them.

Given time, I might see if there is any smart way to use the 8 extra 
registers, but don't count on more than ~10% extra speed in that case. 
We're simply keeping the Opteron SSE unit pretty busy as it is, and 64 
bit pointers doesn't help for less than 4e9 atoms :-)

Cheers,

Erik