The Bad Economics of a 65nm Monolithic Quad-Core

As AMD will soon release Barcelona, it will claim industry first to “native” quad core. Intel on the other hand laughs at the idea as bad economics and foolhardy for the sake of recognition. Producing a monolithic quad core doesn’t make economic sense – according to Intel. Actually even AMD once touted the advantage of a smaller die – ZDNET 2002. But who do you think is correct this time?

The only way to settle this is to try and compute the cost-per-die value of Barcelona against its main rival, the Penryn MCM quad-core. I once claimed that Barcelona costs $174/die to make while it will only cost Intel $53/die to produce a competing product. While it is impossible to prove claims that involve revealing closely guarded secrets such as Fab yield and defect density, it is still possible to demonstrate the impact die size and the native approach has on the economic feasibility of the product.

What we already know:
Barcelona QC die size: 283mm^2
Penryn DC die size: 107mm^2 (214mm^2 for a QC)

To compute gross die per wafer (GDPW), we’ll use the standard formula: {[Pi * (Wafer Diameter/2) ^ 2] / Die Area} - {[pi * Wafer diameter] / (2 * Die Area) ^ 1/2}. Edge exclusion equally ignored. This gives us the result:
AMD at 300mm GDPW = 211; AMD at 200mm GDPW = 79 (quad core dies)
Intel at 300mm GDPW = 622 (dual core dies)

I know that no Fabs are the same, but for the sake of this analysis, let’s assume that defect density at 0.2d/cm2 is equal for Intel and AMD. For Yield we’ll use the Negative Binomial yield model which is commonly used and apply a midrange alpha of 1 for both Fabs, assuming defects follows a gamma distribution. The formula for yield is as follows: [(1+ DieSize*DefectDensity)/alpha]^-alpha
This gives the result:
Wafer yield for Barcelona QC = 63.86%
Wafer yield for Penryn DC = 82.37%
This shows how random defects negatively impacts the bigger die.

Applying the yield against the GDPW this gives the result (GDPW * Yield):
AMD 300mm DPW = 134 ; AMD 200mm DPW = 50 (good QC dies per wafer)
Intel 300mm DPW = 512 (good DC dies per wafer)
Since Intel needs 2 dies to make a QC, each wafer can only yield half: 512/2 = 256 possible QC dies

Assuming once again for argument sake that AMD and Intel has equal Cost Per Wafer. This number is quite complicated as it involves Equipment costs, Depreciation, Direct Labor, materials and Fab utilization so we’ll leave the number equal for both companies to focus only on the impact die size on profitability.

Using the numbers above, in order for Intel to make a quad core MCM Penryn at $53, it has to keep its cost per wafer at $13600 per wafer ($13600/256≈$53). Now if AMD has the same $13600 cost per wafer, its cost per die becomes ≈ $100 at 300mm and ≈$267 at 200mm. Averaging the two if we assume equal loading at Fab30 and Fab36 will give us ≈$184 per die for Barcelona.

The numbers I earlier claimed even comes off as conservative. If we even start to consider the added cost of AMD’s SOI wafer and additional metal layers, the numbers even gets worst for AMD. Again let me reiterate that the numbers above assumes a lot of equal parameters between AMD and Intel which is entirely the opposite in real life. Yet still, the message remains obvious, the impact of a 283mm^2 monolithic quad core on cost per die is massive. The big gap in cost per die between the two competing companies prove that Intel is correct in waiting for 45nm make the same "native" quad core move.

Update: Just in case anyone is wondering, Clovertown's cost is at $76. This is amazing considering Clovertown total die size (143mm^2 * 2) is roughly equal to Barcelona which is 50% more expensive.