Globalfoundries six months on: maybe we could use another foundry

In the discussion of Globalfoundries celebrated spin-off from AMD and subsequently-announced intent to purchase Chartered Semiconductor, one question seems basic: do we really need another foundry right now, and if so, why? Or was the whole exercise really just about getting some big assets off of AMD’s balance sheet? The answer from customers is yet to be heard. Globalfoundries has announced only one customer so far, ST. But the company is clarifying its strategic positioning, and it clearly does not intend to be just one more foundry.

That position, as you might expect, is built on Globalfoundries’ one big customer: AMD’s microprocessor division. This one product line means that Globalfoundries is, by their figures, further up the volume ramp on the 45/40nm node than any other foundry. And that in turn means the good things that happen with increasing volume—decreasing defect density, tighter control of variations and leakage current, more accurate models, and more precise design rules—all get better.

It might seem implausible that Globalfoundries, which has only existed in its present form for six months, could be further along the learning curve than, say, TSMC, which announced 40nm volume production in November, 2008. But that is where AMD comes into the picture. For TSMC, the driving products at 40nm were a couple of big Altera FPGAs and new GPUs from Nvidia and AMD. In none of these did early 40nm production represent a large portion of the customer’s shipments. The products were important, but they did not represent huge numbers of wafers.

In contrast, AMD started ramping 45nm CPU production in Q1 of 2008, and now uses the node widely across its CPU product line. And so yes, it is plausible that Globalfoundries as a primary source of AMD CPUs has run more wafers at 45nm than TSMC has at 45 and 40. In fact, Globalfoundries claims that the difference is huge: according to data derived from International Business Strategies numbers, Globalfoundries claims that in Q1 2009 their shipment rate of 25,000 wafers per quarter compared to a total of about 3,000 wafers per quarter for all other 45 and 40nm foundry processes.

But here things start to get complicated. All those AMD wafers are SoI, not bulk CMOS, using a process jointly developed by AMD and IBM specifically for microprocessors. While many steps in the process are presumably shared with the Common Platform 45nm bulk CMOS process at Chartered and the new bulk CMOS process Globalfoundries is bringing up in Dresden, a custom SoI process is not the same thing as a bulk-CMOS foundry process. So there may be areas in which all that learning curve doesn’t help. This helps explain why Globalfoundries is not planning to start risk production of their bulk 40/45nm process until mid-2010.

Yet clearly there are benefits to having a huge volume running on a very similar process. And those benefits aren’t limited to just 45nm. AMD is already in advanced development on its 32nm SoI process, which Globalfoundries plans to offer as a foundry process starting with risk production at the same time as the 45nm bulk CMOS process, in mid-2010. More significantly for most potential customers, high-performance and low-power 28nm bulk CMOS processes—basically a 10 percent shrink from the 32nm SoI process, redesigned for bulk silicon—will launch risk production in Q4 2010 and Q1 2011, respectively.

All of these 32- and 28nm processes will use the high-k/metal-gate stack jointly developed by AMD and IBM. Unlike most HKMG formulae, this process uses a gate-first fabrication sequence. Globalfoundries claims that gate-first gives them fewer mask steps, higher density, and the ability to directly scale from 32nm to 28nm. From the customer’s perspective, the company says, this means a 5-10 percent smaller die for a given design. Supporting that die-size claim, Globalfoundries is reporting a 0.120 µm² SRAM bit cell and a 113nm contacted gate pitch in the 28nm process, both of which it claims to be the smallest figures so far claimed for a foundry process at this geometry. In addition, the gate-first process will have a design-rule set similar to what the customer will have used before in a 40nm Si-ON process, without some of the new rules introduced to deal with the complexities of gate-last sequences.

So is there a business plan here? It certainly appears from the schedules, the reported figures, and very likely from the yield numbers Globalfoundries is showing to prospective customers that the foundry can in fact lever its production work on AMD CPUs to stay ahead of the rest of the foundry industry on the learning curve. No other foundry will have a customer doing state-of-the-art wafers at anything like AMD volumes. The only other game in that league is Intel. Will prospective early-adopter customers value that learning-curve advantage enough to consider changing foundries? If it translates into higher initial yield and lower risk of design respins, the answer is probably yes.

The next question is the integration with Chartered Semiconductor. In principle, there should be enough similarities between the AMD 45nm SoI process and the Common Platform 45nm bulk process—both of which after all have a common IBM heritage–to make rapid integration of the two companies work, even though the cultural differences are profound. That integration would bring into the game pieces that Globalfoundries absolutely needs: customer-support infrastructure, presence in Asia, an IP infrastructure, an established commercial design flow, and—let us not forget—an established customer base. A foundry that can get all these things working for it will be able to make room for itself at the front table in the industry. But it won’t be a trivial undertaking.