opteron VS Itanium 2

Mark Hahn hahn at physics.mcmaster.ca
Thu Oct 30 10:51:41 EST 2003

> > For bandwidth/memory intensive codes, I think the Opteron is a clear
> > winner in a dual processor configuration because of its dual channel
> > to memory design.  Stream triad bandwidth during SMP operation is
> > ~50% more than a one processor test.  Both the dual Pentium 4 and Itanium 
> > 2 share their memory bus and split (with some loss) the bandwidth in 
> > dual mode. 

this is particularly bad on "high-end" machines.  for instance, several 
machines have 4 it2's on a single FSB.  there's a reason that specfprate
scales so much better on 1/2/4-way opterons than on 1/2/4-way it2's.

don't even get me started about those old profusion-chipset 8-way
PIII machines that Intel pushed for a while...

> This is true as long as you are using an applicaiton where one process has its own
> memory area. If you would have 2 processes and shared memory the Opt, would
> behave like a small NUMA machine and a process will get a penalty for accessing
> another process (processors) memory segment.

huh?  sharing data behaves pretty much the same on opteron systems
(broadcast-based coherency) as on shared-FSB (snoopy) systems.  it's not
at all clear yet whether opterons are higher latency in the case where 
you have *often*written* shared data.

it is perfectly clear that shared/snoopy buses don't scale, and neither does
pure broadcast coherency.  I figure that both Intel and AMD will be adding
some sort of directory support in future machines.  if they bother, that is - 
the market for many-way SMP is definitely not huge, at least not in the 
mass-market sense.

regards, mark hahn.

Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf

More information about the Beowulf mailing list