AMD's Die Yields are only 27%
These EXTREMELY POOR yields are easily corroborated using the publicly available information from AMD as the number of CPUs they have shipped for the past quarter - 7,700,000 - and the number of wafers that AMD professes in their two fabrication facilities - courtesy of Blow Hard Dan - the Ban Ban Monica Man - and he wouldn't lie or try to deceive us - now would he?
For starters, AMD claims to have shipped 7,700,000 CPUs - DudWrongs, AthWipers, AthWiper eXPloders.
Blow Hard Dan tells us that AMD is running 6,000 Wafer Starts per Week (WSPW) through MegaFlop 30 and claims DregsDen (MegaFlop 35) is running at 75% capacity - and that capacity is 5,000 or 6,000 WSPW at Dregsden. Somehow, Blow Hard Dan computes that the actual capacity of AMD was 0.75 x 11,000 = 12,000 WSPW !!
That must be Blow Hard Dan using AMD QuantiMath !!
A better assumption is that AMD's DregsDen fab ran at 80% of capacity in Q3 and that the "new capacity" is just 5,000 WSPW - down 1000 from AMD's prior claims of 6,000 WSPW at Dregsden.
So, 6,000 WSPS (MegaFlop 30) and 0.8*5,000 WSPW (MegaFlop 35) comes to 10,000 TOTAL Wafer Starts per Week for AMD's two MegaFlop Fabs.
Since AMD claims they sold 7,700,000 CPUs in both Q2 and Q3 of this year, and no major "inventory" has built up, AMD's production & sales have remained relatively constant over the past quarter at least. Assuming 13 weeks per calendar quarter (4*13 - 52 weeks), AMD must have processed 13*10,000 = 130,000 wafers (total) during the past quarter at both their wafer fabs.
Let us also assume that AMD had a line yield of 93% (93 good wafers OUT for every 100 wafers started) due to various process issues such as broken and misprocessed wafers - yielding 120,900 actual wafers OUT from these two wafer fabs.
Simple division gives us an overall average Good Die per wafer out as 7,700,000/120,900 = 63.7 die/wafer OUT = functionally good die/wafer.
We now need to calculate the total POSSIBLE number of die on those wafers - the number of complete DIE SITES/Wafer.
AMD claims 50% of their sales last quarter were DudWrongs - die size 104 sq. mm.) and 50% were AthWipers. These AthWipers included 120 sq. mm. regular AthWipers (die size = 120 sq. mm) and approximately 10% AthWiper eXPloders (including AthWiper 4's - the same PaloMeatHead die as the AthWiper eXPloder. The 10% is my estimate. These later die are 129 sq. mm. in size.
Computing an "average die size" is done simply by taking 50% of 104 + 40% of 120 and 10% of 129 (0.5*104 + 0.4*120 + 0.1*129) = 112.9 sq. mm. "in the aggregate" AMD executives love to use this expression - "in the aggregate" !!
Since both wafer fabs use 200 mm wafers, and the area of on 200 mm wafer is 31,415.9 sq. mm, the total possible god die per wafer "in the aggregate" would be 31,415.9/112.9 = 278.
However, many die sites are incomplete because they intersect the edges of the wafer - and typically only 85% of the Total possible die sites are in fact COMPLETE DIE SITES - which would imply that 0.85* 278 = 236.5 are TOTAL COMPLETE DIE SITEs - "in the aggregate" - on these wafers.
We already calculated that AMD produces - on average - 63.7 functionally good die/wafer.
Hence, AMD's die yields can be easily calculated as 63.7/236.5 -= 0.2693175009135 = 26.9%.
Let's be generous and call it 27% !!!
Hence, the available information from AMD - coupled with just a few common sense assumptions - results in the conclusion that AMD is yielding only 27% of the AthWiper/DudWrong die on each processed CPU wafer being functional and meeting their specifications for shipments to customers.
This is an APPALLINGLY LOW number and accounts for AMD losing money despite AMD having a big die size advantage over Intel - whereas Intel's IABG division was quite profitable this past quarter - Q3, 2001. |
