Every phone and tablet chip designer on earth arrived at 28nm inside the same eighteen months, and TSMC, which only just put the node into volume production this month, already has more demand booked than a ramp this young can fill. Qualcomm’s Snapdragon S4 is lining up here, Nvidia’s next Tegra, the FPGA houses, the GPUs, all at once. 28nm is where the mobile industry decided to live, and TSMC is finding out what it means to be the landlord for a market growing this fast.
The thing that makes 28nm matter is under the gate. TSMC finally moved to a high-k dielectric and a metal gate, four years after Intel did it at 45nm, and it took the harder route to get there. Gate-last means building the metal gate after the brutal high-temperature steps that would otherwise wreck it, a more complex flow than the gate-first approach the IBM-and-Samsung camp bet on. Gate-first looked cheaper on the slide. In the fab it has been fighting threshold-voltage control and yield, GlobalFoundries worst of all, stuck debugging two different gate-first processes from AMD and IBM at the same time, and TSMC’s gate-last bet is starting to look like the right call rather than the stubborn one.
28nm is not one process, it is a family. TSMC ships four flavors: a low-cost version that keeps the old polysilicon gate, and high-k metal gate versions, high-performance and low-power and a mobile-tuned one, for the chips that need the leakage win. The tell is what shipped first. Qualcomm’s Snapdragon S4, the marquee mobile part, rides the polysilicon low-power flavor, because that one cleared qualification first; the HKMG mobile process is the step the flagships move to next. A mobile SoC lives or dies on idle current, because a phone spends most of its life with the screen off doing almost nothing, and the leakage you spend in that state is battery you never get back. After the gap that 40nm carried with its old-style gate, this is the node where TSMC stops handing power to Intel’s transistor for free.
While TSMC ships its first high-k metal gate planar transistor, Intel just announced it is leaving planar behind entirely. The 22nm Tri-Gate transistor it detailed in May stands the channel up into a fin the gate wraps on three sides, the biggest change to the device since the industry settled on silicon, and it goes into Ivy Bridge next year. TSMC has nothing like it in production and will not for years. For the first time the structural lead is not just a calendar gap, it is a different transistor, and the whole foundry industry now has a FinFET problem with a clock that started ticking in May.
The customer list tells you who 28nm is really for. The volume is not PC chips anymore, it is application processors and basebands and the first real wave of tablet silicon, riding a smartphone market that has roughly tripled since the original iPhone arrived. And because everyone converged on the same node at the same time, any hiccup at TSMC is now everyone’s problem, where at 40nm a stumble mostly bruised Nvidia alone. The crunch forming right now is that market voting with purchase orders TSMC cannot yet fill: Qualcomm is already lining up GlobalFoundries, Samsung and UMC as second sources rather than wait, and TSMC is pouring better than thirteen billion dollars into a single gigafab, Fab 15, to chase the demand. TSMC already books around half of all foundry revenue, and it means to take an even larger cut of 28nm; the company that owns low-power capacity at this node owns the most important growth market in semiconductors, and the scramble for capacity is the proof.
Perso, I think 28nm sits around longer than the usual node. The shrink after it, 20nm, is going to be expensive, and if it arrives still planar it may not hand back enough power to justify the move for a lot of designs. If that happens, 28nm becomes the node people park on and milk, the cost-optimized default for everything that does not need the bleeding edge. I could be wrong and 20nm could surprise everyone. But the economics of each shrink keep getting worse, and at some point a node stops being a waypoint and turns into a destination.
The leakage gap is closed, the mobile market showed up all at once, and the capacity is spoken for before the flagship parts even ship. That is the good version of TSMC’s year. The bad version is the announcement out of Oregon in May, because a transistor you cannot match for three or four years is not the kind of gap that tunes away like a yield problem does. TSMC spent this node catching Intel on materials. It is going to spend the next two catching Intel on shape, and every customer betting a roadmap on Hsinchu is going to feel each quarter of that chase.