By Mark LaPedus
Faced with huge and unforeseen demand at the 28nm node, leading-edge foundries are scrambling to play catch-up and are boosting their fab capacities at a staggering pace.
But analysts warn that 28nm foundry capacity will be tight throughout 2012, and perhaps into 2013, putting some chipmakers in a pinch. Many blame the 28nm foundry capacity shortfall on a combination of three factors: low yields, lack of installed capacity, and simply underestimating demand. Much of the 28nm demand involves chips for smartphones, tablets and notebooks.
Acknowledging it didn’t have enough capacity in place, Taiwan Semiconductor Manufacturing Co. (TSMC) has responded and now plans to produce 10,000 more wafer starts per month (wspm) at 28nm than originally expected. Rival foundry Samsung Electronics is currently converting two NAND flash lines—one in Korea and another in Austin, Texas—into 28nm foundry capacity. And GlobalFoundries and United Microelectronics Corp. (UMC) are also expanding their 28nm capacities.
“Despite increases in spending by leading foundries, demand for 28nm chips will likely outstrip supply for much of the year,” said G. Dan Hutcheson, chief executive of VLSI Research. “While Apple will get all the 28nm chips it needs, other smartphone makers could be forced to postpone increases in production until more supply becomes available.”
For 28nm, the foundries in total are expected to have a capacity of some 65,000 wspm in 2012, according to Barclays Capital. At 28nm, the foundries are projected to have a total capacity of 300,000 wspm when capacity is fully reached in 2013, according to Barclays. In comparison, total foundry capacity has ranged from 200,000 to 250,000 wspm per node in previous generations, meaning that 28nm could become the largest node in history in terms of volumes.
The 28nm node is emerging as the sweet spot for select, leading-edge products, due to cost and its low-power attributes. Some chipmakers could prolong their efforts at 28nm before migrating to the more expensive and difficult 20nm node.
“We are seeing strong demand for 28nm,” said Michael Noonen, senior vice president of worldwide sales and marketing at GlobalFoundries. “Because of the solid value proposition of this versatile platform, we believe 28nm will be a long-lasting node.”
The big question is whether the foundries will overbuild at 28nm in 2012, causing the segment to fall into an overcapacity mode in 2013. There are some indications the foundries are expanding too fast and customers may be inclined to double order to garner enough wafers.
28nm shortages emerge
Shortages of 28nm foundry capacity, and to a lesser degree, 32nm technology, have been noticeable for months—and for good reason. There is typically an early learning curve for yields at the beginning of each process node.
But the market recently went into panic mode, when Qualcomm disclosed it could not meet demand for its 28nm cell-phone chips throughout 2012 due to unforeseen demand. The shortages involve 28nm parts based on a conventional polysilicon gate stack and not high-k. Initially, Qualcomm opted to use polysilicon at 28nm for cost reasons. In contrast, the FPGA houses elected to use high-k to get more performance.
To meet demand, Qualcomm plans to engage with multiple foundries to make up for the 28nm shortfall. Besides using its key 28nm foundry, TSMC, Qualcomm is now doing business with GlobalFoundries, Samsung and UMC.
Besides Qualcomm, Altera and Nvidia say 28nm capacity is tight. But two others, Advanced Micro Devices (AMD) and Xilinx, insist there is ample 28nm capacity.
Some chip makers reacted quicker to meet 28nm demand. Others weren’t so lucky. In a recent conference call, Jen-Hsun Huang, president and chief executive of Nvidia, said the company will not be able to meet demand for its new 28nm graphics chips throughout 2012. TSMC’s yields are “fabulous” for the device, Huang said. “We under-planned the supply of 28nm. We need to fix that in the future,’’ he said.
Chipmakers impacted by the 28nm shortfall may have overlooked another trend: Less leading-edge capacity is being built at each node. Prior to the 90nm node, many foundry vendors competed at the leading edge.
But as fab and process R&D costs continue to soar, the foundry business has experienced a dramatic shakeout. Fewer foundry vendors—as well as chipmakers—can afford to compete at the leading edge. And leading-edge foundries have become more conservative in capital spending. “The investments to go to the next node are not easy to justify,” said Subramani Kengeri, head of advanced technology architecture at GlobalFoundries.
At 130nm, a fab cost about $1.45 billion, process R&D costs were $250 million, and design costs were $15 million, he said. But at 32nm, a fab runs $4.85 billion, process R&D costs are $900 million, and design costs are $100 million.
“The foundries are being very careful not to overbuild,” said Joanne Itow, an analyst with Semico Research. The net result is that “capacity was tighter at 45nm/40nm compared to 65nm. Capacity at 32nm/28nm is tighter than at previous process generations.”
Today, the foundries are responding to 28nm demand by adding capacity and improving their yields, according to Itow. “28nm capacity will be very tight in Q3,” she said. “The foundries should be able to meet customer demand for 28nm by Q4.”
In 2013 there are questions about what will happen. Some believe the foundries will go into overcapacity mode as demand subsides. Others say demand will remain strong, particularly for 28nm. Mike Splinter, chairman and CEO of Applied Materials said: 28nm “is going to be a long node. We think it will be strong in 2012. It will be strong in 2013.”
Foundries react to demand
During a recent conference call, Morris Chang, chairman and chief executive of TSMC, said the company is unable to meet 28nm demand and vowed it will catch up with customer orders by the first quarter of 2013. TSMC has not been struggling with its yields, but rather it did not have enough capacity in place to meet demand, he said.
In response, TSMC plans to expand its total 28nm foundry capacity to 350,000 to 400,000 wafers in 2012. That’s 10,000 wspm more than originally planned at 28nm.
TSMC is currently producing 28nm wafers in two 300mm fabs. This includes Fab 14 (phase 4 and 5) in Hsinchu. It is also making 28nm devices in Fab 15 (phase 1 and 2), located in Taichung. Another portion in Fab 15 (phase 3 and 4) moved into 28nm production in April. By year’s end, these two phases will churn out 50,000 wspm.
For months, GlobalFoundries has been in 32nm production in Fab 1 in Dresden, Germany. The company has begun ramping up 28nm production in Fab 1, a 300mm plant with a total capacity of 80,000 wspm. In addition, GlobalFoundries soon will move into 32nm/28nm production in Fab 8, its new 300mm fab in New York. Fab 8 is capable of making 60,000 wspm.
Meanwhile, Taiwan’s UMC has begun ramping up 28nm production in 12A (phase 3 and 4), located in Tainan, Taiwan, said Christian Gregor Dieseldorff, an analyst with SEMI.
And not to be outdone, Samsung’s S1 fab in Korea is ramping up 32nm/28nm logic processes for foundry customers, including Apple. SEMI’s Dieseldorff also said Samsung is converting its Line 14 in Giheung, Korea from NAND to 28nm logic capacity.
Samsung, according to Barclays Capital, also is converting a NAND line in Austin, Texas, to 28nm foundry capacity. That line will produce 10,000 wspm a month for Qualcomm, and another 10,000 wspm for other customers, according to the firm.