The chip inside the iPhone 6 that shipped this month is built in Taiwan, and that one sentence is most of the story of TSMC’s 20nm node. Apple moved the A8 off Samsung and onto TSMC’s newest process, the single biggest design win in the foundry’s history, and it did it on a node almost nobody else wants to touch. 20nm got TSMC the most valuable customer in silicon. It may also be the shortest-lived process the company ever ships.
20nm is a genuine density shrink over 28nm, more transistors in less area, and on the power-and-speed axis that actually sells chips it is close to a wash. The transistor is still planar, the same flat structure the industry has used since the beginning, and planar at 20nm is running out of room to drop voltage without leaking. You get the smaller, cheaper-per-transistor die. You do not get the jump in performance-per-watt that makes a phone designer want to move. For a market that parked on 28nm precisely because it sips power, a shrink that does not improve power is a hard sell.
At 20nm the immersion lithography that arrived at 40nm finally can’t resolve the critical layers in a single exposure. The layer gets split across two masks and two passes, double patterning, and now those two exposures have to align to within a few nanometers of each other or the whole layer prints as scrap. All that overhead, the extra masks and the alignment grief, to manufacture a node that gives none of it back in power. The cost-per-transistor curve the entire industry runs on is flattening here for the first time, and a flattening curve is an existential problem for a business built on that curve bending down.
TSMC’s 16nm FinFET arrives next year. It takes this exact 20nm wiring and bolts a three-dimensional transistor on top of it, the fin structure Intel has shipped since 2012, and that is where the power win 20nm withholds finally shows up. Which turns 20nm into a bridge people sprint across. Apple is on it because the A8 schedule could not wait for FinFET, and plenty of other designers are just holding at 28nm until 16FF is ready rather than stopping on a planar node that costs more and barely moves the needle.
Intel is a full structural generation ahead and already ramping 14nm. Samsung has its own 14nm FinFET coming next year and has licensed it to GlobalFoundries, so for the first time two foundries will offer a FinFET at roughly the moment TSMC offers its first. TSMC still books around half of all foundry revenue, but the comfortable distance it kept from everyone except Intel just evaporated. Whoever yields a FinFET first, and yields it well, takes the 2015 mobile sockets, and Apple spent this year proving it will pick its silicon up and move it to whoever has the better process.
Shrink the geometry, get a chip that is cheaper and better: that bargain held for forty years, and 20nm is where it half-broke. Cheaper per transistor if you squint, not better in any way a phone designer cares about, and not even reliably cheaper once the second mask goes on the bill. The only reason nobody files 20nm under stumble is that it carried the A8 and bought TSMC a year to get FinFET right. So the node is mostly a placeholder. What TSMC actually won here is one customer, and the company that owns the door keeps that customer only until 16nm proves it can hold him. Apple just walked off Samsung and onto a Taiwanese fab because the process was better, and he will do it again the second somebody else yields a cleaner FinFET. Apple did not so much hand TSMC a trophy as rent it the socket for twelve months, and the lease renews only if 16nm comes good before Samsung’s FinFET does.