Rob said:
Many thanks Cap'n - I sort of ruled that out, being a new mb and a
ver.2. It turns out that one stick of memory in B1 works - quite why
will be the question for today. But that's certainly handy to know - thanks.
Rob
S939 is not "dual channel", and AMD is the first to admit this,
as they don't use the term in their technical documents. (This
wasn't apparent to me at first, but after reading their docs a
few times, it finally sunk in
B1 sits on the lower 64 bits of the bus. If just B1 is populated,
the CPU talks to memory in 64 bit mode.
If, in addition to B1, a matching DIMM is added to A1, the bus
operates in 128 bit mode. The same command is given to A1 as
is given to B1. In a sense, the two DIMMs are "concatenated" -
another term for this mode of operation, is "Uber-DIMM". A1
cannot be populated, unless there is a DIMM in B1 first. Similarly,
three DIMMs won't work, because once you decide to run in 128 bit
mode, a third unpartnered DIMM won't work.
Now, to make some comparisons, in order of increasing flexibility,
the Intel 875/865 dual channel implementation, supports operation with
three DIMMs. Intel calls this "virtual single channel mode", because
the two channels can be joined end-to-end to make one long logical
channel. The AMD design does not allow this, so in 64 bit mode, only
slot B1 and B2 work. With AMD, you cannot have "lonely DIMMs" sitting
in A1 and A2, without a partner in B1 and B2 first. B1 and B2 sit
on the "primary bus" in a sense.
With Intel, the two channels carry equal significance, so there
is no preferred channel. And, if you mismatch the internal
architecture of the DIMMs on purpose (say, insert a 128MB,
a 256MB, a 512MB, and a 1GB DIMM in the four slots), all four
slots can be used, but they look to the system like a four slot
single channel bus.
The Nforce2 (for AthlonXP) adds to this, the relaxed requirement
that so-called dual channel operation takes place, as long as
there are equivalent memory locations available on each channel.
Say, for example, that you have 256MB+256MB on one channel, and
a 1GB DIMM on the other channel. For the first 512MB of
each channel, the Northbridge alternates between the two
channels. For the upper 512MB of the channel with the 1GB
DIMM in it, the Northbridge can only operate in single
channel mode (as there is no opportunity to alternate
back and forth between channels). I have actually tested
this with a specially modified version of memtest86, and the
memory bandwidth changes from ~1.4GB/sec to ~1GB/sec, as
you pass from the "dual channel" area of memory, to the
"single channel" area of memory.
I believe somewhere along the line, Intel may have added the
ability to relax the channel matching requirements, so that
a similar feature to the Nforce2 is possible. On the 915/925,
the Intel memory guide says you only need to match the total
memory in each channel, in order to enjoy dual channel
operation. That is still not quite as flexible as the Nforce2,
as the Nforce2 can support single channel and dual channel
operation at the same time. With the 915/925, Intel still
operates solely in single channel or in dual channel mode,
and cannot mix the two.
So, there are a number of subtle differences between
processor families and between chipsets. I expect some
of the differences are due to who owns which patents.
AMD's implementation is undoubtedly also intended to
reduce the size of test benches required to verify
the operation of the memory controller, allowing them
to get to market faster.
So, to summarize, with AMD you can use 1, 2, or 4 sticks.
In the one stick case, you can use B1 or B2.
In the two stick case, operation can be single channel
(B1+B2) or dual channel (B1+A1). With the current Intel
dual channel desktop boards, you can use virtually any
configuration, starting with any slot, and get some kind
of working config out of it.
(Note: There are plenty of brain dead memory controllers
out there, so this is just a sampling to illustrate
some differences. Not all Intel designs are as clever,
like the four bank limited 845PE on the P4PE, for example.
Every company has some dumb ideas they would like to
forget, like the 512MB limit on the Intel 815 - something
that prevented me from buying a board with that chip on
it.)
Paul