The A9 in the iPhone 6s that went on sale this week exists in two versions built in two different countries, and that fact is the most interesting thing about TSMC’s first FinFET. Apple split the order, half to TSMC’s 16nm and half to Samsung’s 14nm, hedging its most important chip across two foundries because both of them just shipped their first three-dimensional transistor at the same moment. TSMC finally has the transistor it has been chasing since Intel left it behind. So does its closest competitor, and Apple made them share the marquee.
16FF is the node that finally delivers what 20nm could not. It keeps the same back-end wiring, the same expensive double-patterned metal layers, and swaps the flat planar transistor for a fin standing up off the wafer with the gate wrapped around three faces of it. That geometry hands the gate far better control of the channel, which lets you run at lower voltage without the thing leaking, which is the performance-per-watt jump the planar shrink withheld. Same wires, new transistor, and suddenly the node is worth moving to.
This is TSMC drawing level on shape for the first time in years. Intel has been shipping fins since the Tri-Gate announcement that rattled the foundry industry in 2011, and the structural gap was the one thing TSMC could not close by tuning a process. 16FF closes it. Intel still has a calendar lead and a denser 14nm, but the era where it ran a fundamentally different transistor than everyone else is over. For a customer choosing where to build, the decision drops back to yield, capacity, and price, which is exactly the ground TSMC prefers to fight on, where it holds close to half the foundry market and the deepest capacity in the business.
Notice that TSMC calls this 16nm and Samsung calls its competing process 14nm, and that neither number describes a sixteen- or fourteen-nanometer anything you could point to on the die. The metal pitch on both sits closer to the 20nm generation they share a back-end with. The node name stopped being a measurement a while ago and became a marketing position, a smaller integer for the slide. From here the number on the box and the dimensions on the wafer drift further apart every generation, and anyone ranking foundries on the name alone is ranking press releases.
The dual-sourcing is its own signal. Apple does not split its flagship silicon across two fabs for fun, it does it to keep leverage and to guarantee supply for a launch that ships in numbers no single foundry ramp is comfortable absorbing. The two A9 variants are not identical, the transistors come off different recipes, and the early teardowns are already poking at whether one runs cooler or lasts longer than the other. Whatever that comparison settles on, the lesson for TSMC is blunt: it earned half of the most important socket in mobile, and it has to hold the other foundry off to get the rest.
16FF is the node where TSMC stopped chasing and started racing on even terms, and it should feel like a bigger win than it does. Drawing level is not the same as pulling ahead, and the company it drew level with is sitting in the other half of Apple’s order to prove it. The structural deficit is gone. What replaced it is a straight fight on yield and capacity against an opponent with deep pockets and its own fabs, over a customer who has already shown, twice now, that loyalty lasts exactly as long as the process lead does. 10nm is next, and Apple is watching which of them gets there clean.