Reports from multiple foundries suggest that the industry has had trouble ramping 28nm production, despite early bullish promises and repeated assurances. The situation is something of a high-tech 'I told you so.' As we've previously covered, each new process node has been an increasingly difficult shift as manufacturers have had to ramp new production technologies to build smaller parts. After TSMC's well-publicized problems ramping 40nm, it seemed unlikely that any foundry would be able to painlessly shift down to the 28nm node.
Both GlobalFoundries and TSMC countered these rumors by detailing aggressive product ramps and, in TSMC's
case, a separate 28nm process built using traditional SiON rather than HKMG. GlobalFoundries
has been shipping 32nm HKMG parts since AMD launched Llano earlier this year, but wasn't able to increase production quickly enough to match the admittedly steep ramp AMD was hoping for. When we spoke to GlobalFoundries, the company pointed out Seifert's comments during the Q&A session, in which he stated that the 32nm ramp of Llano—the first commercial HKMG CPU, and the first 32nm chip from AMD—had essentially matched Brazos' ramp. That chip was built on a well-established 40nm process.
Note that wafer shipments is very different from wafer revenue. The newer process technologies produce the bulk of overall revenues
GloFo admits that the HKMG ramp for Llano was difficult but notes that the experience at 32nm has allowed it to ramp 28nm at the same rate. AMD's comments imply that it expects 32nm yields to continue to improve, and that it will launch Llano's successor, Trinity, in Q1 of 2012.
As for TSMC, it recently reported that its 28nm process was in volume production with multiple customers already onboard. Growing economic uncertainty has dragged at the company's overall sales; TSMC's Q3 revenue fell 5.1 percent year-on-year, though its capacity utilization of 93 percent remained excellent. It's been suggested that both the foundry and its partners are leery of over-committing to the new node—partners who got burned at 40nm are wary of moving to 28nm too quickly, while TSMC doesn't want to aggressively ramp a process technology if the demand for it isn't going to materialize as scheduled. All of this is likely to impact AMD's HD 7000 product family launch; it's doubtful that the cards will ship for significant revenue in 2011, despite previous statements implying otherwise from AMD.
Gartner vice president Bob Johnson expects both foundries to focus on improving wafer yields if capacity demand ramps more slowly than expected. With Intel committed to moving Atom to 22nm within a year and the major foundries now running very nearly a full node behind Intel, both manufacturers need to ensure 28nm volumes ramp extremely well. Given the repeated problems we've seen at recent nodes, it may be time for the foundries to revisit their own methods of implementing new technologies. If they don't, they risk falling permanently behind.