When I first heard about AMD's decision to make a native quad-core processor--that is a single piece of silicon comprising all four cores--I thought it was an excellent idea. I began to think of the advantages four-way HyperTransport has over the shared bus architecture of Intel's design. I also saw the significance in the larger shared cache, especially when all four cores are working closely on some particular task, and I didn't really see any downside. However, the more I think about it, the more I am realizing how much better Intel's solution is.
Clovertown is a single-package solution comprising two dual-core Core 2 processors within. It uses a shared bus architecture and is somewhat slower in off-chip communication, but because the die is so much smaller, they can create a larger shared cache for each dual-core, thereby making up for much of that loss.
AMD is not at the same place in manufacturing prowess that Intel is. It has been reported that Intel's yields are also very high, typically 90% or higher with some reports saying 95% or higher. AMD's yields have never (to my knowledge) been publicly disclosed, though I'm going to make a probably very incorrect assumption that they're around 80% for current processors. Even assuming they were 95% for current processors, we know that the number of potential defects goes up by the square of the surface area, so as we move from K8's current 183 mm^2 dual-core 65 nm 512 KB L2 processors, we would see a significant degredation in potential yield due to an additional 100 mm^2 per die alone. That also does not include the potential loss of high-binning parts due to even small defects not making the chip unusable, but rather making them less clockable at higher speeds.
We now know that Barcelona's die will be 283 mm^2, and that Clovertowns are comprised of two 143 mm^2 dual-core dies. Using my Wafer program, we find that AMD could fit a maximum of about 212 Barcelonas per 300 mm wafer, while Intel could fit about 430 dual-core Core 2 processors on a single 300 mm wafer. Now, assuming a 90% yield for Intel, that's about 388 dies per wafer, or 194 Clovertowns. In order for AMD to make as many Barcelonas per 300 mm wafer, its yield would have to be 92%, and by all accounts that's not something AMD will be able to do 1) based on its unconfirmed history of lower yields, 2) based on the natural increase in defect rate due to the larger die size and 3) that defects of any kind which do not render the chip inoperable will render it operable only at slower speeds.
AMD is facing several realities here with its native quad-core design, and it would seem that it does not have the manufacturing prowess to pull it off. Some at Intel even admitted recently that the company would have a hard time doing a die that large with high yields. Note also that Intel's Itanium at its largest on an 180 nm process technology was about 473 mm^2, quite a bit larger. I believe that also might explain why Intel targeted the highest-end markets with Itanium--because those processors sold for so much that very low yields would still be profitable.
As we stand today, with a pricing war going on between Intel and AMD, this native quad-core does not seem to provide enough financial incentive for AMD to have gone with it, even with its potentially much higher performance and scalability.
I see no two ways about it; I think AMD has made a significant mistake in going with the native quad-core design. Had it gone with an approach similar to Intel's, the company could've already had single-package quad-core parts out the door, and their yields would've been greater, their processing power would've only been slightly slower, and they could've been more price competitive.
As it stands, I just don't see the advantages of a native quad-core design. However, I'd very much like to hear your opinions.
Please post your thoughts below.
Source:Clovertown versus Barcelona