Santa Clara — At the GlobalFoundries Technology Conference yesterday, GlobalFoundries executives spoke at length about the company's roadmap and prospects. There was much rejoicing about shipments of the foundry firm's first 32-nm, high-k metal gate (HKMG) chips—otherwise known as AMD Fusion A-series processors, or Llano—but the event really centered on manufacturing at 28 nm and smaller geometries.
First things first, GlobalFoundries revealed that its 28-nm HKMG process is "fully enabled and ready to ramp," with ramping scheduled for 2012 at its fabs in Dresden, Germany and Malta, New York. A "lead 28nm HKMG product" has already taped out (i.e. the chip design is complete and is about to be manufactured), and GlobalFoundries has managed to produce a functional 28-nm HKMG test chip based on an ARM Cortex-A9 core. On that subject, GlobalFoundries said it expects the High Performance Plus version of its 28-nm HKMG process to enable ARM Cortex-A9 processors clocked as high as 3GHz. Now there's something to make Intel sweat, if only a little.
In more general terms, compared to its existing 40-nm process, GlobalFoundries claims its 28-nm process will enable up to 40% greater performance and a reduction in switching power as great as 40% at "nominal operating voltages." With "overdrive," you can look forward to a 50% speed increase and a 50% switching power drop. The company expects to have a leg up on the competition, as well; it hinted that other independent foundries are struggling to get high-k metal gate processes operational, whereas GlobalFoundries has already shipped "tens of thousands" of 32-nm HKMG chips for AMD's Fusion APUs.
Next up is the 20-nm HKMG process, which will enter production in 2013. That process, described by GlobalFoundries as a "full node shrink" from 28 nm, is said to enable a 50% reduction in die area and a performance increase of 35%. GlobalFoundries will use a gate-last approach, as opposed to the gate-first technique it's using at 28 nm. While gate first purportedly enables die size reductions of 10-20% over gate first at 28 nm, GlobalFoundries noted that such density benefits no longer apply at 20 nm. The company doesn't expect to support silicon-on-insulator at 20 nm, either.
Looking further ahead, GlobalFoundries is drawing up plans for 16- and 14-nm fab technologies. I understand the company expects to use extreme ultraviolet lithography and non-planar transistor structures—possibly of the tri-gate variety, like Intel's 3D transistors, at those geometries. GlobalFoundries also said it's shooting for 450-mm wafer sizes for processes smaller than 20 nm. Larger wafers will allow the foundry to squeeze more chips onto each wafer, allowing higher yields and lower costs. We were told that 40,000-45,000 450-mm wafers have a similar die yield to 100,000 300-mm wafers. Costs per die will be lower, too, on the order of 20-25%.
Now, as you may know, Intel plans to move to tri-gate transistors next year—quite a bit sooner than GlobalFoundries. When asked whether that schedule would put the foundry company at a competitive disadvantage, GlobalFoundries R&D chief Gregg Bartlett offered an interesting answer. In a nutshell, he suggested that tri-gate technology is currently immature. While Intel only has its own chips to worry about, GlobalFoundries has to offer processes that can be easily implemented by its many customers—most of whom are not AMD.
Speaking of which, GlobalFoundries provided an update on its overall business. As it turns out, the company now has over 150 customers (yes, that's one-hundred-and-fifty), a lot of whom apparently build system-on-a-chip devices and networking products. Its portfolio of 6578 worldwide patents is also worthy of note, as are its capital expenditures, which totaled $8 billion across 2010 and 2011. Most impressively, the company says demand for its manufacturing capacity currently outstrips supply—in other words, it has to expand rapidly to satisfy its customers.
That expansion involves GlobalFoundries' new fab in upstate New York, which is purportedly ahead of schedule. Production is set to begin there next summer. GlobalFoundries also plans to build a fab in Abu Dhabi—somewhere near the airport, we were told—but a precise timeline hasn't been settled upon yet. (The company says its schedule will depend on ramp planning in Dresden and New York as well as market conditions.) Those fabs will complement the foundry firm's facilities in Singapore, which it inherited after the acquisition of Chartered Semiconductor.
After attending the GlobalFoundries Technology Conference, it's pretty clear that GlobalFoundries has rapidly evolved into much more than AMD's spun-off manufacturing arm. GlobalFoundries seems to be growing rapidly, and its CEO talks of customers pining for an alternative to the current market leader (that would be TSMC). With what seems like a steady flow of cash from its co-owner ATIC, which belongs to the Abu Dhabi government, GlobalFoundries' future may be a bright one indeed.