Re: [Swarm-Support] 64-bit processors

[Marcus G. Daniels]

Steve Railsback wrote:

> So...any idea how to estimate whether a Swarm model is 
> memory-intensive enough to give a 64-bit machine a significant 
> advantage? Lots of agents?  Are method calls important?

The 64 bit capability itself is probably a disadvantage for speed, all 
things being equal.  Having integers and pointers bigger means more 
cache space will be wasted.   However, Opterons have bigger caches and 
more registers so that probably makes up for that problem somewhat.  The 
reason the Opteron/Athlon 64s are good for memory intensive stuff (like 
most agent based models), is that memory latency is lower with 
Hypertransport compared to a Northbridge architecture (as with Intel), 
and bandwidth is higher. 

I would expect that unless a profile demonstrated a simulation had all 
of its bottlenecks in floating point math, that a Athlon 64 3800 or 
FX-53 would be one of the fastest machines available right now on par 
with a 3.4 Ghz Pentium (but the expensive Extreme Edition 32 bit Pentium 
4 would probably be a bit faster due to the large cache).

Roughly, the more agents and non-trivial environmental mechanisms at 
work, the more I'd expect the AMD systems to have an edge over Pentium 
4s and Xeons.  But you should have expectations of differences of less 
than 20%.  The new EM64T Intel systems are clocked up to 3.6 Ghz and 
have a good lead in floating math performance, so there is no clear 
overall leader.   I think low-end Athlon 64 systems are the best value 
right now, though.