[Beowulf] Re: Beowulf Digest, Vol 10, Issue 16

Mark Hahn hahn at physics.mcmaster.ca
Tue Dec 14 09:05:25 PST 2004

> Back to the original thought, which was recycling the PC2100
> memory from the Tyan S2466N boards.  The consensus seems to be that
> registered ECC recycled from the S2466N systems will not work in
> boards designed for unbuffered ECC, which means that both
> the 754 and 939 are out.

I believe the relation is the other way: you cannot use unbuffered
dimms in a board which assumes buffered dimms.  that is, the dependecy
is based on what fanout the drivers can handle: unbuffered dimms 
have more devices hanging off the same signals.  I don't think there's
a difference in pinout.

note that sidedness of the dimm also matters (some dimm vendors 
call double-sided dimms "dual-rank".

> Very unclear to me what, if anything is gained by using 
> registered ECC vs. unbuffered ECC on a smallish system.  On a big

1 clock latency (actually, I think it's a major cycle, so for ddr,
it's the equivalent of two transfers) overhead for registered dimms
(theoretically, there are also buffered dimms which have pass-through
tranceivers which mitigate the drive/fanout problem, but don't 
impose a whole cycle latency.  but I think no one does that anymore.)

> be a plus.  But if the system only holds 1Gb of RAM in one
> or two memory slots the unbuffered memory should be slightly
> faster, and with the ECC enabled, just as reliable.  Correct?

that's my understanding...

More information about the Beowulf mailing list