AMD, GlobalFoundries, and the Intel gap
Gate not closing
When AMD spin-off GlobalFoundries broke ground on its fab in upstate New York last year, the chip manufacturer boasted it was "closing the gap" on Intel. "We were a year behind Intel at the 45nm node, and that difference will be cut significantly at the 32nm generation," said vice president of manufacturing systems technology Tom Sonderman. "By 22nm, there will be no difference. It will be in the noise level."
But a year on, it would appear that the gap is widening. Intel introduced its first 32nm processors this past January, with the arrival of its "Westmere" Xeons, and GlobalFoundries now tells us it's scheduled to deliver 32nm chips by the middle of next year. "The 32nm SOI process is aligned to AMD’s requirements, as they are our only customer for this process at this time," the company tells us. "We have started early production and they have publicly committed that they are sampling now and plan on delivering products in [the first half of 2011]."
The company is also developing a 28nm process — a "half-node" advance from 32nm — and this is slated to deliver chips next year as well. "For the 28nm process — aimed at both high performance and lower power devices — we will tape out products by the end of the year and start production in early 2011," says a company spokesman. "The timing of when you’ll see products is customer-dependent but it’s possible you could see them in [the first half of 2011]."
According to Carl Wintgens, senior process analyst with research outfit UBM Techinsights, the Intel gap is widening in part because the competition is struggling not only with the process-shrink, but also with the move to a high-k metal gate setup, which replaces the old silicon dioxide dielectric with a new high-capacitance metal. Adopted by Intel in late 2007, this setup limits electron leakage from the gate into the pipe that connects a processor transistor's source and drain. And that makes for a transistor that consumes less power and emits less heat.
"Now we're at the end of July, and no other manufacturer has released their 32nm technology. Only Intel has," Wintgens says. "The reason is that most manufacturers will be introducing, for the first time, their high-k metal gate technology at 32nm. It's quite complicated and it can introduce delays."
Intel's high-k setup predates its 32nm process. It debuted at 45nm.
What's more, Intel's competitors are taking a different approach to the high-k metal gate. Intel uses a "gate-last" approach, in which the transistor gate is formed after the source and the drain, whereas GlobalFoundries and other members of the "IBM fab club" are using the traditional gate-first approach. Wintgens argues that the with high-k metal, the gate-first approach can be problematic.
"When you form the source and drain, you have to use a high-temperature process, and if the gate is made of metal, then the electrical properties of that gate will be effected by that process. That's why it's better to make the metal gate after the source and grain. This introduces some complications in the process, but in the end, you have a more reliable process."
One report indicates that Samsung, another IBM fab club member, is exploring a gate-last approach for the move to 22nm, considered that the gate-first approach is only good for one manufacturing "node" — i.e., just the 32nm generation.
GlobalFoundries will likely be the first to complete a 32nm gate-first process, but it would appear that the arrival of actual chips may slip to the middle of next year — though TSMC may nip in ahead. TSMC has adopted Intel's gate-last approach.
GlobalFoundries's 28nm process, Wintgens argues, is merely an effort to provide the illusion that the gap isn't getting wider. "What they're saying is 'we don't want to look too bad compared to Intel, so we'll introduce a 28nm," he says. "28nm isn't a true node. It's a 'mid-node' or 'half-node.'"
He even questions whether the company's 32nm process will make it to market. GlobalFoundries has not yet discussed detailed plans for its 20nm to 22nm process, but the company tells us it will do so in September at an event in Silicon Valley. One can't help but wonder what it will say about the Intel gap. ®