Latest Athlon 64 product introductions

[Yousuf Khan]

I saw my first Socket 939 board at the same time I saw my
first 940 board - in Feb 1993. At that time the AMD rep
at the demo was predicting general availability of socket
940 chips for April 2003 (which actually happened) and
socket 939 chips for Q3 2003. So far Q3 2004 is just a
few weeks away and no socket 939 chips seem to be in stock
in any stores.

Well, that's some forward thinking on AMD's part, they were already prepared
with a 64-bit processor socket ten years before they released it.

Yousuf Khan

 

[Grumble]

But the first Ed was saying that a Socket 745 required six layer
mobos, while Socket 939 requires only 4 layer.

Just a nit. Socket 754, not 745.

Go ahead, count 'em

http://www.hardspell.com/newsimage/200392613445242585.jpg

 

[Rob Stow]

Well, that's some forward thinking on AMD's part, they were already prepared
with a 64-bit processor socket ten years before they released it.

Yousuf Khan

I meant last year. 2003, not 1993. 2004-1 = 1993
That has got to be the weirdest "typo" I've ever done.
Either that or my math qualifies me for a gov't job.

 

[Yousuf Khan]

Lots of 'em support three. Remember, the 939 is still dual-channel
(right?). Are you suggesting that each channel will only have one
DIMM? Perhaps that's the only way to make it work, but I'm glad I
went with the 940, if so. Still once you've transformed the wires
out to the first DIMM, the traces to the second are pretty much
parallel, thus I still don't understand the issue.

Well, it's a dual-channel in the sense that the processor reads the lower
64-bits from the first DIMM and the upper 64-bits from the second DIMM. I
don't think it actually puts the two chips in separate banks and does bank
interleaving in the traditional dual-channel sense. This is how Pentium 4's
dual-channel works as well.

Yousuf Khan

 

[Yousuf Khan]

...and the 940 doesn't need 'em. I'm not buying that one either.

Well, I don't either, just passing on what's been stated elsewhere.

Perhaps. But why wouldn't they build that into the 940 footprint
too? Leave the desktop with faster memory (down the road, of
course) than the "server" chipset?

That shouldn't be too difficult to accept. The server market is
traditionally more conservative about technological changes than the desktop
market is. So it's likely that once DDR2 has proven itself in the Athlon 64,
the Opteron will require a totally new socket other than S940.

Or perhaps Intel's FB-DIMM dreams will come true, and you won't be requiring
a new socket to move to DDR2, because it will work in existing DDR1 slots.

Yousuf Khan

 

[K Williams]

There is nothing inherent in the Opty that limits motherboard
manufacturers to 4 DIMMs per processor. HP, for example, has
managed 8 in their 4-way systems.

I didn't say there was. However many 940 boards only support four
(mine, fer instance. Why would a 940 board that only supported
four take any more wiring channels than a 939 board that supported
four? For that matter, once you've fanned the signals out to the
memory modules, why would more memory modules require more layers?
Both assuming dual-channel, of course.

 

[K Williams]

Probably because a 940 board will truly be able to support 4 DIMMs
whereas a 939 board will more than likely be limited to two DIMMs
in reality. When was the last time you saw desktop boards
routinely support more than two DIMMs?

Lots of 'em support three. Remember, the 939 is still dual-channel
(right?). Are you suggesting that each channel will only have one
DIMM? Perhaps that's the only way to make it work, but I'm glad I
went with the 940, if so. Still once you've transformed the wires
out to the first DIMM, the traces to the second are pretty much
parallel, thus I still don't understand the issue.

I can only remember this
happening back in the Pentium 1 days. Ever since then, all I've
ever seen is warnings against putting any more than two DIMMs on
any board even if it has space for more.

Remember, It's dual channel, so there would only be two on each
channel (for a total of four).

The only ones that can truly do it these days are the buffered
DIMMs on server boards.

Sure at these speeds, and two per channel.

We're still missing something here.

 

[K Williams]

Well as some people have pointed out in the case of Pentium 4's
Socket 775 -- extra grounding.

....and the 940 doesn't need 'em. I'm not buying that one either.

Then again it's possible that maybe AMD is already preparing for
DDR2, and Socket 939 is a forward looking design preparing for
that day.

Perhaps. But why wouldn't they build that into the 940 footprint
too? Leave the desktop with faster memory (down the road, of
course) than the "server" chipset?

 

[K Williams]

Well, it's a dual-channel in the sense that the processor reads
the lower 64-bits from the first DIMM and the upper 64-bits from
the second DIMM. I don't think it actually puts the two chips in
separate banks and does bank interleaving in the traditional
dual-channel sense. This is how Pentium 4's dual-channel works as
well.

How is this different than interleaving? An address and command is
sent to both channels and each fetch (in the burst) comes from
alternating channels. There are still two independent sets of
wires for the two channels, no? Otherwise I don't see what
dual-channel buys, electrically (note that I haven't looked that
closely).

 

[Yousuf Khan]

How is this different than interleaving? An address and command is
sent to both channels and each fetch (in the burst) comes from
alternating channels. There are still two independent sets of
wires for the two channels, no? Otherwise I don't see what
dual-channel buys, electrically (note that I haven't looked that
closely).

Well, I guess it's just that in this case, both DIMMs are really just
extensions of one another inside the same bank. One row-column address is
asserted that activates both DIMMs. Whereas in the bank-interleaved
dual-channel, there has to two separate, but simultaneous, invocations of
rows and columns, one for each bank since they each DIMM resides in separate
banks.

Yousuf Khan

 

[K Williams]

Well, I guess it's just that in this case, both DIMMs are really
just extensions of one another inside the same bank. One
row-column address is asserted that activates both DIMMs. Whereas
in the bank-interleaved dual-channel, there has to two separate,
but simultaneous, invocations of rows and columns, one for each
bank since they each DIMM resides in separate banks.

I don't see the difference, except in some picky details. In reality
that's all they are in an "interleaved" memory system too. Both
leaves get the same address and they respond *interleaving* the
data. On the classical systems, the interleaved banks would share
the data bus, as well. Here they're completely separate, since the
bus bandwidth is completely used without "interleaving". ...still
not really anything new. It's not like the different controllers
are fetching from different places in memory.

 

[Nate Edel]

Probably because a 940 board will truly be able to support 4 DIMMs whereas a
939 board will more than likely be limited to two DIMMs in reality. When was
the last time you saw desktop boards routinely support more than two DIMMs?

I built a lot Slot-1 systems with desktop boards and 3 DIMM slots. A fair
number of them didn't take 256mb DIMMs, and I upgraded a lot of those to 3x
128mb DIMMs.

For that matter, I've got a BX-based slot-1 server board with 4x 256mb
unbuffered, which seems perfectly stable.

I can only remember this happening back in the Pentium 1 days.

I recall it being more part of the transition from PC-100/133 to DDR.

Ever since then, all I've ever seen is warnings against putting any more
than two DIMMs on any board even if it has space for more. The only ones
that can truly do it these days are the buffered DIMMs on server boards.

On modern DDR systems, that seems correct, but back in the good old days of
the BX chipset...

 

[Rob Stow]

Full-sized ATX desktop boards almost always have 3 or 4 DIMMs.
You seldom see only two unless the board is an ITX or mini-ATX.
If the DIMMs are non-registered, getting DIMMs to work in the
third and fourth slots can be tricky, particularly if they
are large DIMMs, but simple configurations like
(3 or 4 ) x (256 MB or 512 MB) usually work.

 

[Rob Stow]

I meant last year. 2003, not 1993. 2004-1 = 1993
That has got to be the weirdest "typo" I've ever done.
Either that or my math qualifies me for a gov't job.

And I just noticed that when I saved my W2K hardware profile
a few months ago I gave it the label that now shows up in
my boot menu ... "Saved 1 April 1994".

 

[GSV Three Minds in a Can]

Bitstring <[email protected]>, from the wonderful

And I just noticed that when I saved my W2K hardware profile
a few months ago I gave it the label that now shows up in
my boot menu ... "Saved 1 April 1994".

I guess it must be age + wishful thinking. I've only managed to be off
by a couple of years so far. Start worrying when you find yourself
typing 1904.

 

[Yousuf Khan]

And I just noticed that when I saved my W2K hardware profile
a few months ago I gave it the label that now shows up in
my boot menu ... "Saved 1 April 1994".

Time to jump into the old time machine.

Yousuf Khan