RDRAM vs SDRAM redux

[Robert G. Brown]

Dear List Humans,

Life continues to get more puzzling all the time.  We are working out
final configurations for a mixed purchase of P4's and Athlon XP's.  Or
so I thought when I started to review the hardware alternatives this
morning. I'm basically getting ready to update a quote from three months
ago but the world has of course changed substantially in the meantime.

The Athlon update was fairly easy.  It looks like the KT266A chipset is
probably the one of choice for a single CPU solution (which I'm inclined
to) and in the meantime 512 MB DDR PC2100 DIMMS are now cheaper than 256
DIMMS were in the first quote.  Also choosing the XP for a single CPU
choice is a no-brainer.

The P4's are much more difficult because there are now SDRAM chipsets.

Does anyone have words of wisdom (or benchmarks!) to offer for the
performance of P4's running e.g. lattice QCD or other numbers,
especially those illustrating differences between code that uses SSE
instructions? I already found

  http://qcdhome.fnal.gov/cluster_design/benchmarks.html

but it is a bit dated (being all of five months old:-) and doesn't
include KT266A and XP OR SDRAM-equipped P4's.

I'm especially interested on what the best choice would be for a P4
intended to do well on memory-intensive code, e.g. Intel 845 (SDRAM but
CPU up to 2 GHz) or 850 (RDRAM but only 1.8 GHz?) or SiS 645 (DDR up to
2 GHz) as there are getting to be a truly dazzling array of
alternatives.

An obvious question is whether or not our lattice QCD folks and/or
quark-gluon plasma folks really need to get the P4's to hedge their bets
at this point.  The benchmark results above at FNAL show the P4 holding
a small (~20%) lead over the Palomino out in the large lattice sizes
likely to be dominated by memory speed.  The stream results for the P4,
especially with SSE instructions, still are much better for the P4 than
(say) the Palomino, but the KT266A suppposedly delivers 20-30% better
DDR performance than the KT266 did (and maybe than the AMD 760 used on
the Tyan Thunder?).  There is also no clear indication on whether using
an SSE compiler with the XP makes a difference -- does the XP support
SSE1 and/or SSE2 instructions?

Sigh.  Any help on these questions would be greatly appreciated.  Also,
if the FNAL folks are listening and have some newer boxes handy, it
would be fabulous of you to update your benchmarks above.

  rgb

-- 
Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu