Extreme Ultraviolet Lithography, or EUV, has been stuck somewhere between a mirage and reality for over a decade, shimmering faintly on the horizon with promises of “4-5 years away.” Meanwhile, other technologies like immersion lithography and multipatterning have been deployed to extend the limits of 193nm lithography far past its original limit. But even the best research teams and scientists can only extend something for so long. Last week, we had the opportunity to tour GlobalFoundries’ Fab 8, in Malta, N.Y., and to sit down and talk about how the company’s roadmaps and timelines look through 2018 and beyond.
The fab space is astonishing to see in person. You know academically how complex these things are, but seeing them is something different.
First, despite being built from the ground up as a 28nm fab back in 2012, the fab has been converted entirely to 14nm, with initial 7nm pilot production also underway. All previous nodes have been moved to other facilities. GlobalFoundries 22FDX platform (that’s the 22nm planar FD-SOI solution), for example, is built in Dresden, at the former AMD foundry now known as Fab 1.
Moving EUV From Wishful Thinking to Reality
There’s been a great deal of talk in the semiconductor industry about when EUV would arrive and what node it would use. GF is working on introducing EUV at 7nm, but it’s only going to see limited use for contacts and vias, not as part of the critical path. One thing we saw while at the foundry was the ongoing installation of two massive ASML NXE:3400B tools to support the company’s future production (GF has cleared space for four machines in total).
One reason GF is moving forward with a limited initial deployment of EUV is because the pellicle has been so difficult to develop. Using extreme ultraviolet light at 13.5nm is so difficult because EUV is absorbed by literally nearly everything. That’s why the inside of the manufacturing system has to maintain vacuum conditions. Right now, there isn’t a good pellicle solution for EUV (the pellicle is a transparent cover that sits over the mask and prevents particles from landing it). Currently, only about 77 percent of EUV light can make it through the pellicle layer, and that’s not enough for full integration into the production line.
Why do you need to keep particles off the mask? Because, without a pellicle, any foreign object on the mask will be printed to the wafer. And while some pellicle solutions have been found, they don’t tend to survive well at the high source power required for full production integration.
EUV has made great strides on multiple issues, but pellicles remain problematic. Image by ASML.
Here’s the (extremely simplified) problem: One of the biggest problem with EUV has been source power. All else equal, higher source power = higher throughput, and a modern foundry lives and dies on its wafers per hour throughput. The long-term goal is to find a pellicle solution that can handle high initial source power (up to 250W) at higher efficiency rate (the current pellicle is ~77 percent efficient, the goal is ~88 percent).
HVM = High Volume Manufacturing.
GlobalFoundries’ solution? Use EUV for contacts and vias while a pellicle solution is being worked on. Since you don’t need one for these areas of the chip, you can increase EUV throughput and reduce cycle time. Going forward, GF will adopt EUV for more critical mask layers. The CTO and SVP of Worldwide R&D at GF, Dr. Gary Patton, has suggested EUV is a functional requirement for 5nm or below. The sheer number of masks required at that point could otherwise make it infeasible for any customer to justify using the technology.
One major question about GlobalFoundries’ future has been how their 7nm solution would compare with that from other companies. The company’s problems with its own version of 14nm, 14XM, were bad enough that it killed that product and licensed IP from Samsung instead. At 7nm, GF is trying again to bring its own process technology to market, and it’s promising some impressive advantages. Customers should be able to deliver the same performance at 60 percent reduced power, or increase performance by up to 40 percent. Unlike its competitors, GlobalFoundries is skipping 10nm altogether and heading straight for 7nm, with an AMD Vega chip designed for machine intelligence workloads apparently serving as a so-called “pipe cleaner” to test the design and its capabilities.
Also: For those of you curious about 12nm and how it compares with 14nm, GF confirmed its 12nm is a refinement and improvement to an existing process with some optimized layouts that offer up to 10 percent performance improvement or a 15 percent density increase. We don’t know how AMD will use 12nm yet for Zen+, but if I had to bet, I’d bet more on performance enhancements and less on sheer density. AMD will have had an eye on that process as a way to close ground compared with Intel’s Coffee Lake.
There are still questions about how GF will pull off the transition, how well EUV will ramp, and what the future holds for the company, but the mood at GlobalFoundries was optimistic on both counts. It’s not surprising, in retrospect, that GF had trouble out of the gate. The company was spun off at a time when its largest (and initially, its only) customer was on life support. It faced huge challenges in ramping up production and earning new customers.
Today, the company seems to be on stronger footing. AMD, which is still GF’s biggest publicly known customer, is itself doing much better. Overall, GloFo seems to have significantly strengthened its own position in the foundry market, but it’ll be at least a year before we know how much 7nm business the company has picked up compared with the competition.