AMD poised to exit 2006 with 55% market share (run rate)

Various news indicate that AMD is ramping up FAB36 and Chartered FAB7 as fast as it can, furthermore, 65nm products will be shipping as early as June 2006. Clear indication that AMD is gunning to take over 50% of the x86 market by the end of 2006.

Keep in mind, as of 4Q05, AMD took 21.4% market share with one single 200mm FAB, the FAB30. What it can do with two additional 300mm FABs is staggering. Previously, it was reported that AMD was ramping FAB36 to 13000 wspm. Now, assuming Chartered FAB7 contributes 5000 wspm for AMD64, we reach a 18000 wspm for the 65nm node. The capacity estimate is done again with the Wafer application from Geek.com, which computes the number of dies off a wafer.

For this, let's use the biggest known AMD chip, the Socket 1207 Rev F dual core Opteron as a reference. At 90nm, the dize size for Rev F dual core Opteron is 220mm^2, at 65nm, the die size is thus 110mm^2. Other dual core chips such as Socket AM2 Athlon 64 X2s are 25% smaller, their die sizes are around 85mm^2. To be conservative, let's use a geometry of 10 mm x 9.5mm for the Wafer program. We get 690 dual core dies off each 300mm wafer.

The annual dual core die output at 18,000 wspm is thus: 18000* 12 * 690 = 149 million. With a very conservative yield of 55%, we get 82 million working dual core CPUs.

For FAB30, let's assume AMD sticks to 90nm and continues to produce the same CPUs for the lower end there, FAB30 thus contributes 50 million units. We reach a total AMD dual core CPU output of 82+50= 132 million per year at the end of 2006.

According to IDC, global PC units will increase 10% to about 230 million (with US PC units growing to 69.5 million).

132 / 230 = 57.4%

Intel has about a dozen production FABs, however, most of Intel's FABs are oudated. Only four Intel production FABs are capable of running at 90nm or 65nm*. To be more specific, Intel has only two 65nm production FABs and two 90nm production FABs. In fact, the $10 billion Itaniums are still stuck at 130nm, while the Xscale mobile chips are still at 180nm. For some chips, Intel has to rely on Taiwanese foundry TSMC. Due to Intel's sixth generation architecture(Bob Colwell, 1995), its CPUs needs minimum of 2x2MB of cache to have acceptable performance, the large cache requires large die area. For instance, the 65nm Presler has a 162mm^2 die size, the Conroe's die size is about 155 mm^2. Intel is planning to build additional FABs. However, since Intel has only $10 billion cash but 100K people to feed, its ability to add capacity depends on making profit or selling stocks, both of which are becoming harder and harder, as its products sit in the lower end of the performance spectrum.

AMD will widen its performance lead as we progress in 2006. Dual core Turion 64 X2 will be unleashed soon at 25 watts, then Socket AM2, Socket AM3, Rev F Opteron, Quad-core Opteron and the K8L. The hounds, the K-10 and K-11 are also on schedule. AMD will soon add a new set of extensions to AMD64 to keep the clone makers busy.

* Intel has two 65nm FABs (FAB12, FAB24-2), two 90 nm FABs (FAB24, FAB11x), seven 130nm FABs, and a bunch of 0.18, 0.25,0.35,0.50,0.70,1.0 micron junkies.

*revised to use a lower estimated yield for the 65nm node.

PS: There is a great secret on Intel's capacity, if you can see it, please post your findings in the comments area.
PPS: since the secret is now exposed, I modified the main text to include Intel capacity analysis.

PPPS: In case you didn't notice, Intel did a shelf registration with SEC today, which stated "Under Intel’s second restated certificate of incorporation (the “certificate of incorporation”), Intel is authorized to issue up to 10 billion shares of common stock". Two weeks before earnings, time running out.