You are using an out of date browser. It may not display this or other websites correctly.
You should upgrade or use an alternative browser.
You should upgrade or use an alternative browser.
E
Ed
Wow and I thought 940 was a lot, IOW what AMD is going to do with all
these extra pins!
Ed
K
Keith R. Williams
Wow and I thought 940 was a lot, IOW what AMD is going to do with all
these extra pins!
Gotta hold the chips down somehow (lotsa grounds and Vdd). ;-)
G
GSV Three Minds in a Can
from the said:Wow and I thought 940 was a lot, IOW what AMD is going to do with all
these extra pins!
Second core, perchance?? Well, you know, all the extra power, ground,
and data, to keep the second core fed and happy. 8>.
R
Richard Krehbiel
Ed said:Wow and I thought 940 was a lot, IOW what AMD is going to do with all
these extra pins!
12 HyperTransport channels!!!!
Nah, probably not.
K
keith
Second core, perchance?? Well, you know, all the extra power, ground,
and data, to keep the second core fed and happy. 8>.
I don't see where the second core would cost any significant I/O.
Power/ground, certainly.
Actually, I'm still amazed at a 940pin package for $hundreds. Pins are
expen$ive (30ish years ago we used 1800 pin modules, at six-figures apiece).
R
RusH
G
Grumble
Ed said:Wow and I thought 940 was a lot, IOW what AMD is going to do with
all these extra pins!
http://www.amdzone.net/modules.php?name=Sections&req=viewarticle&artid=56
http://www.amdzone.net/pics/cpus/dualcore/1stdemo/diagram.jpg
<quote>
Each CPU has a path to the system request interface, and through the
crossbar switch shares the same memory controller, and 3 HyperTransport
links. AMD feels that the single memory controller is able to handle
both CPU cores. They also feel that the HyperTransport bandwidth
provided by three full speed links is more than adequate. AMD documents
reveal that perhaps a 10% drop in performance due to the shared components.
</quote>
Could AMD want each core to have its own memory controller?
N
Never anonymous Bud
Trying to steal the thunder from Arnold said:DDR2 ?
Dual-cores, of course.
G
GSV Three Minds in a Can
from the said:I don't see where the second core would cost any significant I/O.
Power/ground, certainly.
Depends whether you want each core to have it's own memory controller &
HT links I guess.
Actually, I'm still amazed at a 940pin package for $hundreds. Pins are
expen$ive (30ish years ago we used 1800 pin modules, at six-figures apiece).
Economies of scale .. plus technical advances. Heck I remember when 64
pin (DIL) packages for UARTs were expensive because they had to be
ceramic, rather than plastic, and we were contemplating how to get past
-that- barrier. 8>.
R
RusH
Never anonymous Bud said:
of course no ? they are gonna be 940, thats what amd showed last time
Pozdrawiam.
Y
ykhan
GSV Three Minds in a Can said:Second core, perchance?? Well, you know, all the extra power, ground,
and data, to keep the second core fed and happy. 8>.
Well, they said they are going to be able to fit dual-cores within the
existing S940. But with S940, you only have one memory controller
feeding both cores. Maybe this next gen socket will allow for dual
independent memory controllers too? Also likely by that time they'll
be doing DDR2 too.
Yousuf Khan
C
chrisv
Well, they said they are going to be able to fit dual-cores within the
existing S940. But with S940, you only have one memory controller
feeding both cores. Maybe this next gen socket will allow for dual
independent memory controllers too? Also likely by that time they'll
be doing DDR2 too.
You mean a 256-bit-wide memory interface? Wow.
K
keith
Depends whether you want each core to have it's own memory controller &
HT links I guess.
Other than that's not what AMD has been saying, another few hundred pins
isn't enough to do what you propose. A single memory channel would take
more than a hundred I/O.
Economies of scale .. plus technical advances. Heck I remember when 64
pin (DIL) packages for UARTs were expensive because they had to be
ceramic, rather than plastic, and we were contemplating how to get past
-that- barrier. 8>.
Sure. Note that we're still in ceramic. ;-)
K
keith
Well, they said they are going to be able to fit dual-cores within the
existing S940. But with S940, you only have one memory controller
feeding both cores. Maybe this next gen socket will allow for dual
independent memory controllers too?
I don't buy it. Add up the pins. 1207 isn't enough. Also remember the
articles about the "extra" port on the K8's memory switch, just sitting
there ready for another core.
Also likely by that time they'll be doing DDR2 too.
Ok, why more I/O?
R
Robert Redelmeier
keith said:Actually, I'm still amazed at a 940pin package for
$hundreds. Pins are expen$ive
What was the old rule? A dime apiece? So?
These things aren't wirewrapped & potted.
AFAIK, current packaging is like ultra-high precision BGA
(30ish years ago we used 1800 pin modules, at six-figures apiece).
Yeah, but you didn't make Millions of modules per year.
Economies of scale!
-- Robert
Y
Yousuf Khan
chrisv said:You mean a 256-bit-wide memory interface? Wow.
More like two independent 128-bit wide interfaces.
Yousuf Khan
T
Tony Hill
Depends whether you want each core to have it's own memory controller &
HT links I guess.
No particularly good reason for each core to have their own HT links,
that just complicates things for basically no improvement in
performance. Even dedicated memory controllers are unlikely to be
worthwhile, Intel has demonstrated quite clearly with their Xeons that
two CPUs can share a single memory controller with very little loss in
performance vs. AMD's NUMA design.
Now, if they try to slap *4* cores on a single die, now that's another
matter altogether.
Still, my money is on those extra pins being almost entirely made up
of power and grounding pins.
K
Keith R. Williams
What was the old rule? A dime apiece? So?
Purchased, socketed, manufactured and tested, closer to a dollar
apiece.
These things aren't wirewrapped & potted.
;-)
AFAIK, current packaging is like ultra-high precision BGA
X86 processors are still PGA. The marketing model still demands it.
Certainly BGA would be cheaper.
Yeah, but you didn't make Millions of modules per year.
Economies of scale!
No, not millions per year, but a hundred per machine. Of course they
did get better at making them and the cost went down. What killed 'em
was when it dropped to a half-dozen per machine. The economy of scale
went all to hell. ;-)
~1K pins still amazes me. 1K balls, less so.
J
JW
X86 processors are still PGA.
Not entirely.
Ask a Question
Want to reply to this thread or ask your own question?
You'll need to choose a username for the site, which only take a couple of moments. After that, you can post your question and our members will help you out.