By David Lammers
Sometimes, you have to hear it from the horse’s mouth. Everyone has heard about the claims of improved 32nm yields at the GlobalFoundries Dresden fab. At the Common Platform Technology Forum, details emerged about how the GlobalFoundries team boosted yields so dramatically.
Mike Noonen, hired away in January from NXP Semiconductors as the sales and marketing senior vice president at GlobalFoundries, started out the Common Platform event by acknowledging that in the third quarter of last year, low yields were “challenging” the volume production of AMD’s Llano processor.
“We took a variety of steps in the fab and on the management team,” Noonen said at the Santa Clara event Wednesday (March 14). “The result was that Llano yields doubled in a quarter. We are in a very, very aggressive ramp of Llano,” he told more than a thousand attendees.
Noonen said the foundry achieved an 80% increase in 32nm unit shipments from Q3 to Q4.
How did GlobalFoundries’s Dresden yields improve in such a short time?
An executive gave credit to the “great German engineers” at Dresden, with a key boost from Robert Madge, director of design-enabled manufacturing at GlobalFoundries, who was sent over to Dresden early last year to help fix the yield issues.
Rutger Wijburg, hired last August to run the Dresden campus after working at NXP’s foundry operations, brought an intense focus on yield improvements, this source said. Rather than try to accommodate a dozen new foundry products at the same time — which “sent the Dresden operation into shock, hurting everything” — Wijburg preached “Llano, Llano, Llano” to the Dresden team. “German engineering did the rest,” the executive said.
Single-wafer clean tools from Dai Nippon Screen (DNS) were used much more widely, a key factor in the yield-enhancement campaign. “Getting the particles off the wafers with single-wafer cleans was one of the main things they figured out. We also bought a bunch of brightfield inspection tools. And there were some back-end copper issues that got figured out,” he said.
With 32nm SOI yields now at very respectable levels, “AMD is happy again,” the executive said, with the relationship back on a normal customer-supplier basis. If GlobalFoundries can deliver the goods, they will keep AMD’s business, even in the face of competition from TSMC at 28nm bulk and 28nm HKMG.
AMD has its “Trinity” processor coming along this year, and the expectations are for AMD to account for about $1.5B in business for GlobalFoundries this year, up from $800 million in 2011. JoAnne Feeney, a stock analyst at Longbow Research, said AMD has been gaining market share, with a 17 percent share of the notebook processor market.
Feeney said notebook customers pay a lot of attention to battery life, and the AMD notebook processors are equivalent to what Intel-based systems deliver. Consumers don’t know or care if the processor is 32nm SOI or 22nm Tri-gate, she added, as long as the performance, graphics, and, especially, battery life are competitive.
The AMD-based thin notebooks — which Intel has branded as UltraBooks – may be significantly cheaper than the Intel-based UltraBooks, Feeney added. While Intel has outlined specific guidelines for what UltraBooks must be able to do, thereby jacking up the costs, the AMD customers have a relatively free rein and may be able to undercut the UltraBook retail prices by a couple hundred of dollars, she said.
“I am predicting that AMD will gain share this year,” Feeney said at the Common Platform event.
Another interesting point was raised by a senior marketing manager working at GlobalFoundries. While a year ago some major smartphone IC manufacturers were saying they would mainly use a non-high-k gate stack for 28nm applications processors, the advantages of high-k are now swinging more of their product mix toward a high-k solution. The nitrided polySi gate oxide just can’t keep up in terms of power and performance with the more-complex high-k gate stack. And with mass production of 28nm parts slipping somewhat, customers are tilting toward high-k at 28nm and pushing back tapeouts of 20nm parts by three or four quarters, he added.
With several hundred thousand wafer starts of 32nm high-k/metal gate production behind it for Llano production, the company’s painful yield enhancement process will pay dividends at 28nm HKMG, he said.
Another executive at the Common Platform Forum claimed that through all of the foundry’s 2011 challenges, GlobalFoundries customers have been willing to cut the foundry some slack, largely because they don’t want to be totally dependent on TSMC and thereby vulnerable to wafer price increases as 28nm heads into mass production.