Intel 14nm shows up this month in Broadwell, in a low-power laptop chip called Core M, and for the first time in memory it is late. Not catastrophically, a few quarters past where tick-tock said it would be, but late, and that is new. The metronome that has not missed a beat since the mid-2000s just missed one. Intel says it is a one-time stumble on an unusually hard shrink. The transistors are good. The schedule, for once, was not.
The node itself is excellent. It is Intel’s second-generation FinFET, the three-dimensional transistor it shipped first at 22nm, now smaller and more efficient, and on pure density and power it is the best process in the world by a wide margin. That is the thing to hold onto through the delay: even a year behind its own schedule, Intel’s 14nm is ahead of anything the foundries can build. Broadwell runs cool and efficient enough to go fanless in a laptop. Whatever went wrong at 14nm, it was not the transistor.
The delay has a cause worth understanding, because it is the first sign of something larger. Each shrink has been getting harder, and 14nm is where the difficulty curve finally bent faster than Intel’s engineering could. Patterning features this small with the old deep-ultraviolet light means more and more exposures stacked on each other, more steps, more ways for yield to go wrong, and Intel chose to push that approach further rather than bet early on the EUV machines that still are not ready. For now the bet looks fine; 14nm works. The question nobody is pressing hard yet is what happens at the node after this, when the same problem is worse.
Down a tier, the foundries are not standing still. Samsung just built the 64-bit A7 on 28nm and is pushing hard toward its own FinFET, TSMC is right behind on its own 28nm, and both are about to reach the three-dimensional transistor Intel has had since 2012. Their FinFET nodes land next year, branded 14 and 16nm even though the geometry sits closer to Intel’s 22nm than its 14nm, the node numbers having quietly detached from anything physical. The three-year lead Intel enjoyed at 22nm is quietly compressing, not because Intel went backward but because it paused, and a pause is all the foundries need. They have Apple’s volume and Apple’s money funding the chase, and they are closing the distance a year at a time.
Intel still does not have a phone business, still treats the foundry model as beneath it, and now has its first schedule slip in a decade, and the company is filing these as three unrelated facts rather than one pattern. The pattern, if you want to see it, is that the integrated model is a flywheel that only spins while the manufacturing lead is enormous. Shrink the lead and the logic gets shakier: the premium margins depend on being ahead, and the closed fabs depend on the premium margins. Intel is not worried. It has been ahead so long that ahead feels like a property of the company rather than a result it has to keep earning.
One missed beat is not a trend, and it would be foolish to write off the best manufacturer in the world over a single late node. Perso, I expect Intel fixes the cadence at the next node and this 14nm slip becomes a footnote, the year the metronome hiccuped and then recovered. That is the likely case. The uncomfortable case is that the slip is not a hiccup but the first symptom, that the physics has finally gotten ahead of even Intel, and that the company is about to learn the difference between a lead it earned and a lead it assumed. Probably the first. Intel has earned the benefit of the doubt. It is just the first time in years it has needed it.