How hard would it have been to MCM the athlon X2?

[taltamir]

I am curious as to how hard it would have been for AMD to make an MCM 4x offering using two athlon X2 die back when intel first came with a quad core? I am assuming it would have taken them a few monthes at most, probably just a couple of week of r&d. And that it was mismanagement that made them pass it up.
But I could be completely off on that.

So is there anyone here with experience in the field who could answer this question?

 

[myocardia]

While I can't give you a definite answer, I can tell you that it would have been decidedly faster than trying to go monolithic, like they did with Phenom. Of course, it still would have been much slower than the Kentsfield, clock-for-clock, just like the X2's were to the C2D's. At least the Phenom lowered that advantage by a considerable margin, plus it does give them something to "build on" now.

 

[taltamir]

Well, but what about last year. I think that if they went MCM they would have had a quad offering a year ago, or more. And the phenom would have came out maybe a month later for it.

 

[ViRGE]

Without changing the motherboards, the only way to MCM the X2s would be to enable one of the disabled HT links on the X2 die, then have a second X2 die connected to the first via that HT link. You'd be looking at a sort of master-slave design where one die would be the main one and housing the active IMC, while the second die would be operating in NUMA mode. Basically the second core would be slower at most things as a result, which isn't ideal since I can't imagine any OS is prepared for an unbalanced NUMA setup.

 

[jmmtn4aj]

Originally posted by: taltamir
Well, but what about last year. I think that if they went MCM they would have had a quad offering a year ago, or more. And the phenom would have came out maybe a month later for it.

How does that work?

 

[taltamir]

It works the way intel did it... make an MCM offering right away by putting two die on a single module. And a year later make a unified die... and then put two of those on a chip aswell to get the next one up...

AMD instead waited, releasing the X2 much later then the Pentium D. but it annihilated it in performance... but soon after intel finished their true duel core processor which was single die. And soon after started offering quads... while amd again waited for almost a year before countering with a "true" single die quad core... but this time the performance benefit wasn't enough to even compete with the archetictually better core 2.

 

[Aluvus]

Originally posted by: taltamir
It works the way intel did it... make an MCM offering right away by putting two die on a single module. And a year later make a unified die... and then put two of those on a chip aswell to get the next one up...

This is very easy for Intel and very hard for AMD, due to the factors ViRGE noted. The design of K8 or K10 is very different from Netburst or Core. The technical challenges of making an MCM K8 work are far from trivial. AMD would also have to learn how to package an MCM, and buy any additional equipment needed.

AMD instead waited, releasing the X2 much later then the Pentium D. but it annihilated it in performance...

The Athlon 64 X2 launched the same month as the Pentium D. In fact, the press releases (1, 2) are spaced a mere 5 days apart. A dual-core Extreme Edition part had launched earlier, but you couldn't actually buy it.

 

[taltamir]

oh, now THAT is the kind of information I was looking for.

So you say there are archetictural differences between intel and AMD that make an MCM much more difficult for AMD. Ok, thats a very good thing to know. Can anyone give specific examples of what those archetectural differences ARE?

 

[Cookie Monster]

Economies of scale, and various other things when it comes to business makes me believe that such a move for AMD would have been even more costly than what they are doing now.

 

[roblob]

I think the main challenge for making a MCM chip for AMD is that K8 and K10 use an integrated memory controller.

 

[DrMrLordX]

If they had wanted to go MCM, they would have needed to produce single-socket AM2 boards with two sets of memory sockets so as to avoid the "unbalanced NUMA" configuration that ViRGE described above. I think that would have been a fairly good idea considering the fact that AMD had to update their platform and memory controllers to accomodate DDR2 anyway. Sure, there would be no quads for s939 users but compared to what's happening to AMD now, that would not have been such a big problem.