[gmx-users] Re: Opteron: compile with 32-bit SSE

[Erik Lindahl]

Hi,
>
> In addition to my last question on this, just out of curiousity (the
> functioning of cpus still seems very complicated to me): might the 64 
> bit
> code for the opteron then be more powerful on really large systems or 
> in
> double precision, compared to the 32 bit? Do you maybe know a simple 
> text
> that can help understand how this all (register accessing etc) works?

64-bit architectures will allow you to simulate systems where the total 
size of the neighborlists is larger than 2Gb, but that will require 
some (minor) changes in the neighborsearching. Normal systems aren't 
approaching this limit yet.

The only really good references I know of when it comes to register 
renaming are the hardware manuals that you can download from Intel and 
AMD - the locations change now and then, though, so you'll have to go 
to e.g. developer.intel.com and use the search facilities.

As for 32 vs 64 bits it's much easier: the latter have 64-bit integer 
registers and can do 64-bit arithmetics in a single cycle. There are 
very few applications that need this (Gromacs is for instance entirely 
limited by floating-point performance), so the only clear advantage is 
that you can allocate more than 2Gb of memory in a process.

The drawback is that all pointers are 64bit. If you do a lot of pointer 
manipulation (which is bad, but anyway :-) you will waste twice as much 
cache on the pointers. Further, on some architectures like PowerPC you 
have to load a 64bit pointer in two steps to maintain backwards 
compatibility, and that can reduce performance slightly.

Cheers,

Erik