Plans are on hold as industry digests the cost and need for larger wafers.
By Mark LaPedus, Ed Sperling & Katherine Derbyshire
There was a time not very long ago—one process node, in fact—when the economic momentum of Moore’s Law seemed unstoppable with a combination of extreme ultraviolet lithography, larger wafer sizes and a variety of new materials. Shrinking feature sizes is still technically possible, but certainly not with the same promised economic benefits and, at least for the foreseeable future, not with 450mm wafers.
At one time, Intel, TSMC and Samsung were aggressively beating the 450mm drum. Chipmakers wanted, if not demanded, 450mm pilot line fabs in place by 2016, with high-volume manufacturing 450mm plants by 2018. They have since altered course, leaving the next-generation wafer size in limbo. In the past few months, all of the major toolmakers and foundries, except Intel, have stepped back to reassess.
No one is definitively ruling out 450mm, but even Intel agrees that it is on hold until the end of the decade. TSMC is not as vocal as Intel, but the Taiwan foundry giant is still interested in 450mm, as well. Still others believe 450mm fabs could get pushed out to the 2020 to 2025 timeframe. And, as happens with all high-stakes gambles—particularly in the wake of EUV’s perpetual delays—there are bets being placed that 450mm will never happen.
Long-term commitments required
But if chipmakers really want 450mm, they will need to support it sooner rather than later. Equipment makers need a timely decision in order to move forward with their 450mm tool R&D efforts.
“Industry consensus should be reached soon to keep the latter part of the decade intact (for the advent of 450mm fabs),” said Hamid Zarringhalam, executive vice president of Nikon.
At present, Nikon is developing a 450mm lithography tool, based on 193nm immersion. Last year, CNSE/SUNYIT, the Global 450mm Consortium (G450C) and Nikon announced a $350 million partnership to develop next-generation lithography tools for 450mm fabs. Nikon plans to ship the 450mm tool to the G450C by next year. “Our 450mm immersion scanner will ship to Albany on time in April 2015,” Zarringhalam said.
It’s no surprise that Nikon is moving full speed ahead with 450mm. The company’s largest customer is Intel, one of the proponents of 450mm technology.
KLA-Tencor, meanwhile, introduced four new inspection and review systems at Semicon West this month, optimized for bare silicon wafers as well as complex multilayer stacks that for the basis of FinFET and 3D-NAND structures. While the company’s previous generation Surfscan SP3 was and continues to be available for both 300mm and 450mm wafers, the Surfscan SP5 is available for 300mm wafers only. The reason: 450-mm wafers are unlikely to drive volume equipment purchases for several years, by which time the market will be looking for a still newer generation of inspection tools.
Most fab tool companies are in somewhat of a bind. On one hand, they sell equipment to Intel and TSMC. On the other hand, many also sell tools to chipmakers that have no plans to build 450mm fabs, namely the memory giants, such as Micron and possibly Samsung.
“The memory players are saying: ‘We don’t want to go 450mm. We’ll never go 450mm,’ “ said Dave Hemker, senior vice president and chief technology officer at Lam Research. “To me, if that’s the case, that’s almost the deal breaker right there.”
On top of that, toolmakers do not have the resources to fund the development of both leading-edge 300mm and 450mm gear. “If I need a tool set for memory at 300mm, and a logic tool set at 450mm, it’s going to be tough unless someone is paying me to do the development. It’s simple economics when you look at it,” Hemker said.
Hemker also raised the age-old issues with 450mm. The market is limited because there are only a few customers that will buy 450mm gear. “I am investing a whole bunch of money to cut my market size in half,” he said. “The flip side is if I can charge twice as much, I’m fine.”
In any case, Lam Research has been developing 450mm tools for some time. “If you look at the 8- to 12-inch conversion, a whole bunch of the benefits at 12-inch came from things that had nothing to do with the wafer size change. For example, we went to automated systems and FOUPs. The equipment became better in that the yields improved. We also became better at defect management,” he said.
“That’s how we structured our 450mm program. We tried to pick the things to work on first, which we know can be ported back to 300mm. This includes uniformity improvements and cost,” he added.
GlobalFoundries, meanwhile, is on the fence while the economics of 450mm and stacked die are sorted out.
“There are multiple efforts to tie 450mm with 2.5D and 3D,” said Ramakanth Alapati, director of package architecture and customer technology at GlobalFoundries. “A 2D chip on 300mm wafers at 10nm is on par with the cost of a 3D chip. When you get to 7nm and 5nm, 3D is cheaper. And the cost of going to 450mm may actually be more expensive. Do we really want 450mm and 7nm if analog and I/O don’t scale?”
He said that analog and I/O saw one-time improvements by moving to finFETs at 14nm, but that won’t be repeated at 10nm. On top of that, multipatterning at 10nm will increase defectivity, so simply making a wafer larger won’t actually improve the economics of a feature shrink because the yield will decrease.
“We’re looking at multiple options,” Alapati said. “For logic and memory, you can use 2.5D to do logic partitioning. The flatter they are, the better the case for 2.5D. For functionality, we’re looking at logic partitioning in 3D. With bandwidth-limited applications, such as GPUs and MPUs, you need a large amount of memory with short distances.”
He noted that 2.5D will begin showing up in limited volume next year, with a volume ramp in late 2015 and 2016 as yield increases. If the foundries can bring down the defectivity quickly enough, he said the focus for stacked die will shift away from yield (cost) to performance benefits of this packaging technology. And at that point, 450mm could become a very tough sell to the semiconductor manufacturing industry.