Intel 22nm ships this spring in Ivy Bridge with the first three-dimensional transistor ever put into mass production, a fin standing up off the silicon with the gate wrapped around three sides of it. The launch parts are the quad-core third-generation Core chips, a 22nm shrink of last year’s Sandy Bridge. Intel calls the transistor Tri-Gate; the rest of the industry calls it FinFET and does not have it yet. For the second time in five years Intel has taken the single most important transistor change of the generation and shipped it first, alone, while everyone else is still reading the papers.
Below a certain size the gate stops controlling the channel. Current leaks across the device whether it is switching or not. Shrinking the transistor stops buying speed or power, and the flat planar design every chip has used for decades simply runs out of road. The fin is the way around it: stand the channel up and wrap the gate over three sides, taking back the control the flat layout had been losing. Intel built that on the high-k foundation it laid at 45nm and refined at 32nm, and built it years ahead of anyone else. The foundries know FinFET is coming and know they cannot skip it. They are about three years from shipping their own.
TSMC is shipping 28nm, a planar node, a genuinely good one, and its customers are thrilled with it; it is also a flat transistor at the exact moment Intel has moved into the third dimension. Samsung sits in the same spot, building planar parts for Apple while betting on the gate-first recipes Intel already walked away from. TSMC is talking about a 16nm FinFET around 2014 to 2015, Samsung and GlobalFoundries a 14nm in the same window, all of it still on the drawing board while Intel ships fins this spring. Three years is the gap, and in this business three years is forever. Intel looked at all of them, saw a generation or more of daylight, and reached the conclusion it always reaches: the integrated model wins because manufacturing is the moat, and right now the moat has never been wider.
A three-year transistor lead, gross margins above sixty percent, and a tick-tock cadence that has not missed in years. The balance sheet could buy most of its rivals outright, if Intel ever felt the urge. The question of opening the fabs to outside customers comes up again, the way it does every couple of years, and Intel swats it down the way it always has: why would the company three years ahead of every foundry rent that lead to the fabless firms those foundries exist to serve. The crown jewels stay locked in the vault. The whole strategy fits in one line, that Intel makes the best chips because it makes them on the best process, and at 22nm in 2012 nobody can argue with it. This is the peak, even if no one inside the building would call it that.
Phones are the threat Intel is not taking seriously enough. It finally has silicon in a few this year, an Atom called Medfield, but it is a 32nm afterthought in a handful of handsets, nowhere near the 22nm crown jewels, while ARM and the foundries quietly build an entire parallel industry that does not need Intel at all. Underneath that sits a quieter problem, almost invisible from inside 2012: the next shrinks. 14nm is promised for next year, 10nm penciled in for 2015, and each one gets harder than the last as the physics turns mean exactly where the features turn small. Intel has always out-engineered the physics. The assumption that it always will is doing an enormous amount of silent work in every roadmap on the wall, including the one with those two dates on it.
None of this is a real problem at 22nm, and betting against a company this far in front would be foolish. Perso, the FinFET lead alone should hold Intel out front for years; you cannot conjure a three-year process gap out of nothing, and the foundries are not close. What the lead has done, though, is make Intel certain, and certainty is a strange thing to carry into the exact stretch of the roadmap where the physics gets hardest. Intel shipped the future again this spring, ahead of everyone, and it has every reason to believe it always will. The only thing it is not doing in 2012 is asking what happens the year the shrink finally fights back.