Intel 45nm chips started shipping this month with something nobody else has managed to put into mass production: a transistor whose gate is built from a hafnium-based high-k material and metal instead of the silicon dioxide and polysilicon the industry has leaned on for forty years. The launch parts are the Core 2 Extreme QX9650 on the desktop and a wall of Xeons for servers, codenamed Penryn, die-shrunk Core 2 chips with bigger caches and the new SSE4 instructions. The products are an evolution. The transistor underneath them is the story. Intel just solved the leakage problem the whole industry has been slamming into, and it solved it first, in volume, the way Intel solves most things.
The gate is the part of the transistor that switches it on and off, and the insulating layer under it has been getting thinner with every node until it is only a handful of atoms thick, so thin that current leaks straight through it and the chip wastes power doing nothing. High-k metal gate replaces that failing insulator with a thicker layer of a material that does the same job without leaking, and swaps the polysilicon gate for metal, a different metal for the n-type and p-type transistors, to make it work. Intel claims more than a tenfold cut in gate leakage and around twenty percent more drive current out of the change. Gordon Moore called it the biggest change to the transistor since the late 1960s. Intel built it gate-last, forming the metal gate near the end of the process, the harder path that happens to yield better.
The quieter decision at 45nm is the one Intel did not make: it refused to change its lithography. The rest of the leading edge is moving to immersion scanners at this node, but Intel extended plain 193nm dry patterning by imposing strict, regular design rules and reusing most of its 65nm tooling, taking only a small single-digit increase in cost to double the density. It is the conservative-genius move, change the transistor materials, which nobody else can do yet, while holding the riskier tool change for 32nm. The contrast with the rivals is sharp. The IBM alliance, a dozen-odd firms pooling process R&D, AMD and Samsung and Chartered and Freescale and STMicroelectronics among them, is going gate-first high-k on immersion 45nm and aiming for 2008 into 2009, betting that forming the metal gate early is cheaper to integrate. TSMC is bringing up its own 45nm immersion process for next year but holding high-k back, leaning on the lithography shrink alone while it works the materials separately. Same node name, several different bets, and Intel’s gate-last, dry-litho version is the only one already shipping.
This is tick-tock working exactly as designed. Intel alternates years: a tick shrinks the existing design onto a new process, a tock puts a new design on the matured process, and the metronome has made Intel the most reliable manufacturer in the history of the industry. 45nm is a tick, the shrink, and it arrives on schedule with a transistor breakthrough bolted on. While the pure-play foundries are still building out 65nm for other people’s chips, Intel is a full node ahead and inventing the transistor everyone else will have to copy.
The lead is not close. AMD, Intel’s only real rival in x86 and holding under a quarter of the market and sliding, is a process generation behind, stuck at 65nm and fighting its own fires this year, and it does not expect 45nm parts until the middle of 2008. The foundries that build everyone else’s chips, TSMC chief among them, are a year or two back and working on problems Intel already solved. Intel designs the chip and builds the chip and owns the factory, the integrated model, and right now that model looks unbeatable because the hardest part of making a modern processor is the manufacturing, and nobody manufactures like Intel. The company sets the pace and everyone else measures themselves against it.
There is a question Intel does not bother taking seriously, which is whether it should build chips for anyone but itself. The foundries are growing, fabless companies are multiplying, and some people think manufacturing will eventually separate from design across the whole industry. Intel’s view is that this is a lower-margin business beneath a company that makes the best processors in the world. Why rent your crown-jewel factories to outsiders when your own chips earn far more per wafer. From inside a two-year process lead, the logic is hard to argue with.
The thing about a lead this large is that it makes you certain, and certainty is expensive. Intel is sure its integrated model is permanent and just as sure the foundries are a tier below. Perso, from where things stand at the end of 2007, it is hard to see what changes that. Intel has the best transistors, the best schedule, the best margins, and a manufacturing arm a decade ahead of its design rivals. Maybe the only real risk to a machine this good is the day it stops running on time, and there is no sign of that day. The metronome has not missed a beat in years.