There were a number of charges for the first quarter that did have an overall negative impact, but in the long run can be viewed as positive. The first charges were those related to the acquisition of server company SeaMicro. This company specializes in high density, low powered units, which rely on a unique fabric interconnect technology that is very efficient at what it does. Currently these servers rely on Intel Atom based processors, but we can assume that they will transition to Brazos 2.0 and Trinity based units. This transition is a win/win for both SeaMicro and AMD. SeaMicro will have near unlimited access to these processors at a lower price, achieve better 64 bit performance over Atom, and also have a robust integrated graphics portion with OpenCL support. AMD obviously benefits here by moving more low power parts into the server market and by providing unique products to the market that simply are unmatched by other outfits. This is an aggressive acquisition for AMD, and hopefully one that will pay off for both companies.
The large and unpleasant charge that we see this quarter is that of the wafer agreement change with GLOBALFOUNDRIES. AMD has now divested itself fully from GF, and that is reflected in the negative results for this quarter. AMD owns no part of GF, and that is most likely a good thing. Losses incurred by GF are no longer reflected in AMD’s quarterly results. This also has resulted in slightly more flexibility for AMD in negotiating wafer agreements with GF. Now AMD can bring up the specter of cheaper production with TSMC and UMC to potentially allow for better deals with GF. Unfortunately, these new deals will not match the one-time sweetheart deal that AMD was able to get while still being a joint owner of GF. The upside again is that GF and ATIC can absorb any potential losses and missteps that are endemic to the foundry industry, and AMD will be unaffected by this.
There has been some confusion to the second part of the major loss incurred by the $703 million charge credited towards GLOBALFOUNDRIES (made up of both cash and the rest of AMD’s stake in the company). Some have taken this as a negative and that it is essentially AMD buying themselves out of the exclusive agreement to use GF’s 28 nm process. My understanding is that it is quite the opposite. In the conference call Rory Read explicitly stated that AMD will be producing 28 nm parts through GF in 2H 2012, and those parts will be released to market in early 2013. AMD in fact bought exclusivity on GF’s 28 nm process. Currently TSMC is maxed out on 28 nm production and cannot offer a significant amount of space to any new customer. So for AMD to try to produce CPUs on TSMC’s process, they would have to redesign for a gate first manufacturing technique, and they would have to cut down their graphic’s orders to be able to get any CPU based wafers out. This is obviously not a good solution for AMD, so they are tapping GF to make these next generation 28 nm parts based on the Piledriver CPU architecture and the GCN graphics architecture.
This is a good trade off. AMD is taking a charge now, but the potential advantages of having exclusive use of GF’s 28 nm process is quite big. Unlike GF’s current 32 nm SOI (Silicon on Insulator
) process, there will be plenty of customers lining up to use the bulk 28 nm HKMG line. AMD is moving away from SOI wafers for now, but that does not necessarily mean that they will not adopt fully depleted SOI in the future. The current 32 nm production does use partially depleted SOI wafers, but the jump to 22 nm might in fact include fully depleted SOI. This is still up in the air, and we will not know what AMD’s and GF’s plans for FDSOI will be until well into next year.
GF has really hit its stride with 32 nm, and this will have a direct effect on 28 nm production later this year. Getting rid of SOI wafers will lower the complexity of manufacturing, and decrease the cost per wafer. While GF has been behind in getting 28 nm out the door, they do look to be firmly on track for ramping production aggressively for the second half of this year. Once production starts up, AMD will be lowering orders on 32 nm, and GF will then be able to start swapping over that line space to 28 nm.
Trinity in the flesh!
Going to 28 nm production is a very good step for AMD. Obviously they are adjusting their designs to run on bulk silicon rather than SOI. 28 nm HKMG is still chock full of material advances which will give switching performance that should be on par with the current 32 nm SOI process, and transistor leakage again should be very comparable between the two products. What AMD is really gaining here is the smaller die size with much the same power and thermal characteristics of the slightly larger 32 nm SOI process. We can assume around a 15% to 20% decrease in die size as compared to the older process. Using Trinity as an example, we see that at 32 nm it is around 240 mm squared in size. If we go with the upper limit of a potential shrink of 20% we have a die that is now 192 nm squared. On a 300 mm wafer we would see approximately 253 gross dies per wafer (assuming a 2mm edge exclusion and counting incomplete dies) for the 32 nm version, but we jump to 313 gross die. This is a pretty hefty jump in both gross die and (assuming similar yields) good die. We also must factor in the lower cost of bulk silicon wafers vs. SOI. The GCN architecture also appears more efficient per mm squared than VLIW4, so the assumption is that the die space needed for the graphics portion will not grow significantly from what is currently being seen in Trinity.
These factors will help AMD be more competitive with Intel and their 22 nm process that the current Ivy Bridge chips are being manufactured on. AMD will still be at a disadvantage, but it will not be nearly as bad as when relying on 32 nm SOI process technology exclusively. If AMD can effectively utilize half node jumps for their CPU products, they can cut off anywhere from 6 to 8 months of Intel’s overall process lead. When we consider that Intel is around 16 months ahead of where AMD is now, that is a significant win for AMD. So basically when Intel makes the leap to the next process node (in this case 22 nm) AMD will make the jump to 28 nm around 7 to 8 months later. This is a much better scenario for AMD rather than waiting another 18 months to go to 22 nm. Eventually they will hit 22 nm about 9 to 10 months before Intel goes to 14 nm. This again helps keep AMD more competitive with its larger rival.