EduardoS (no@spam.com) on 12/31/11 wrote:
---------------------------
>For a time AMD had >40% 4P x86 server market, maybe more than 50%, Intel have a
>specific die for this market so a new mask should pay itself, of course, for AMD
>it's better to earn more money without having to pay an extra mask, but since they
>don't have a real competitor against Westmare-EX earn without paying isn't an option anymore.

Those days are long over and today Intel has a much bigger share of the server market AND sells at much higher margins, so they have a lot more money to spend on new masks.

>BTW, Thuban was almost a server specific die, latter it was marketed in the high-end
>desktop market where the volume is even lower than servers, falling later to the
>mid-end, where margin for such a big die is pretty low, and it lost a lot of ground
>on server space, what if AMD had a better server die?
>
>Of course, using the desktop die on the server is better, but if it is not possible a server specific die pays itself.
>

I don't know if Thuban wasn't meant for the desktop originally, but it's hardly a high-end low-volume part in the sense you're speaking. It was a $200-300 part and the only response AMD had against Intel's i7 Nehalems. I would really hope that AMD sold far more of them as Phenom II X6 than as Opterons.

The real role the die played in the server space was in 8 and 12-core Magny Cours chips, which goes against your point of designing new dies for the server space. Here Thuban was MCMed (sort of ironically given AMD's earlier stance on selling MCMs to get more cores) to pretty good effect - note that this also involved using two Thubans for an 8 core part over even making a new Deneb-like die to be conjoined. This supports that AMD doesn't see it as worthwhile to invest in saving this level of die space on server parts and I'm sure we'll see the same with 12 core Bulldozer Opterons (and even 8 core MCMs!). It can even make more economically sense if the parts with fused off cores let you salvage bad yields, or if the truly lower core dies yield worse for it.

Maybe it's true that sometimes you need to invest a lot now to save more in the long term but AMD's hands are tied, they just don't have the money to invest anymore (they're in substantial debt at the moment). The single die approach probably hurts desktop manufacturing costs more than server costs, since Zambezi has some server features fused off. But it can't really be helped at this point. Actually I would say that AMD is better off NOT throwing more money on new dies for products that won't have a lot of demand in the first place.