Supermicro PIIIDM3 and PIIIDR3 motherboards

Alan Scheinine scheinin at crs4.it
Fri Jun 23 04:05:00 PDT 2000


   I am new to this mailing list, but I did try to read the Archives
of this year before posting this question.

   I want to buy (have assembled from boards) a Intel-based PCs
that will give me good bandwidth with Myrinet.  In particular,
I would like a 64-bit PCI bus, so the field is narrowed to just
a few motherboards.  More specifically, there are two motherboards
from Supermicro for which I would like to hear some realworld
experience: PIIIDM3 (memory PC100) and PIIIDR3 (memory Rambus).

   On 25 April 2000 Eric billings submitted a table that showed
the PIIIDME (same as PIIIDM3 but without built-in SCSI) as having
slow STREAM results, slower than the chipset 440BX.  Keith Underwood
replied, "That's odd.  Tim Carlson had posted STREAM benchmarks
earlier for that board that were much better (2x or more).

   In looking over the archive, I did not come across a follow-up
article.  Eric Billings wrote, "We are continuing to test 840-based
motherboards as they become available."  So perhaps it is appropriate
to revisit the question.

   Brian Haymore wrote that he returned 5 or 7 of PIIIDME boards and
runs them without with memory ECC turned off.  Aside from the ECC problem,
which another contributor said has been fixed, could Eric Billings comment
on the speed of running programs in comparison to a 440BX chipset?  Does
the interleaving of PC100 memory access on the 840 chipset give a visible
speedup?  What I mean to say in general, in addition to STREAM tests, does
anyone have comparisons running actual memory-intensive programs?

   With regard to Rambus, I did not see any mailing list messages
concerning the Supermicro PIIIDR3, has anyone tried it?  I read
that for the 820 chipset the memory translator that allows the use of
PC100 adds latency.  Because the 840 chipset is so new (with regard
to its market share), I have not come across comments concerning a
possible analogous problem for the 840 chipset.  If Rambus really
does give a significant speed-up, it might be worth the cost; especially
since the cost of Rambus is gradually declining.

Alan Scheinine  Email: scheinin at crs4.it




More information about the Beowulf mailing list