[gmx-users] Re: scaling issues in GROMACS

[Erik Lindahl]

Hi,

A couple of things:

1. The early versions of gcc for the G5 had a bug that affected CPU 
detection. Make sure the logfile says that Altivec loops are used.
2. By default you compile in 64-bit mode on Opteron. I have written 
AMD64 SSE assembly loops, but they are not yet in the distribution. You 
will get essentially the same performance by compiling in 32-bit mode 
with SSE. (64-bit mode without SSE is significantly slower).
3. I have new assembly loops for itanium2 that are screamingly fast, 
but this is for the next version too :-)

Without assembly loops on ia64, I would guess the best performance 
solution is to use fortran in the inner loops and disable the software 
sqrt, but your milage will vary...

If you're thinking of investing in machines, ia64 will be the fastest 
by far once the new loops are out, but also pretty expensive. G5's and 
Opterons should yield about the same bang per buck when both use 
assembly loops - let me know what you're seeing...

Cheers,

Erik