GLOBALFOUNDRIES' CTO State-of-the-Roadmaps

At SEMICON I got to sit down with Gary Patton, the CTO of GlobalFoundries. They have a dual roadmap, with both 14nm FinFET (licensed from Samsung) going to 7nm (internally developed), and with the other track being FD-SOI (licensed originally from ST Microelectronics at 28nm) at 22nm and 12nm (internally developed). For more details on the dual roadmap see my post GlobalFoundries' Dual Roadmap.

History

Let me start with a brief history of GlobalFoundries (GF from now on), in case you don't know it.

Jerry Sanders, the founder and longtime CEO of AMD, made the blunt statement that "real men have fabs" in the early 1990s, meaning that having control of semiconductor manufacturing and tying it into product strategy was crucial for a serious semiconductor company. But that was before the foundry and fabless revolution really took off. I say revolution, but really it was more of a gradual transition. Semiconductor companies discovered that they could no longer afford the expensive luxury of their own fabs (and their own process) and so they went fab-lite, milking older processes in their existing fabs, but switching to foundries for leading-edge designs for which they no longer had an in-house manufacturing capability.

Ironically, it was "real men have fabs" AMD who first abandoned manufacturing totally. Of course, they had several fabs, most notably Fab 1 in Dresden Germany where their microprocessors were manufactured. They didn't do the fab-lite thing, they did something more radical. They sold their entire manufacturing operation to ATIC, which is the investment arm of the Abu Dhabi government in 2009. The new company was called GlobalFoundries. AMD became completely fabless (and GF became the real men?). And AMD solved the "real men have fabs" thing by becoming the first major seiconductor company, in Lisa Su, to have a female CEO. "Real Women don't have fabs."

GF immediately started construction on a new fab in Malta (upstate New York, not the Mediterranean island). Since GF was spun out from AMD in March 2009, and construction started in July 2009, I assume some level of the work had already started under AMD's watch.

Less than a year later, in January 2010, GF acquired Chartered Semiconductor, a pure-play foundry, third biggest at the time, based in Singapore and with several fabs there.

Then, in 2014, they acquired IBM Microelectronics, the semiconductor manufacturing arm of IBM, although IBM did retain some semiconductor R&D. Gary told me soon after the sale that he actually had the choice of staying at IBM and running the R&D there or moving to GF and taking the CTO role, which is obviously what he chose.

So today GF has the following fabs:

• Fab 1 in Dresden, running 22FDX and older processes, 300mm, will run 12FDX
• Fabs 2-6 are 200mm fabs in Singapore (were Chartered)
• Fab 7 has been upgraded to 300mm (was Chartered)
• Fab 8 in Malta NY running 14nm, 300mm, will run 7nm
• Fab 9 in Burlington VT (was IBM) 200mm running mostly RF I think
• Fab 10 in East Fishkill NY (was IBM) 300mm, running silicon for IBM servers in partially depleted SOI, and probably other stuff by now
• Fab 11 in Chengdu China (under construction, JV) 300mm, will run FD-SOI at first

GF had AMD as a somewhat captive customer as part of the spin-out, but struggled to get a lot of new customers as the new kid in town with no track record as a foundry. Having been part of the team that spun Compass Design Automation out of VLSI Technology, I know how much more complex it is than it seems at the start, even to the participants. There are all sorts of hidden dependencies. For example, AMD presumably had no team whose job was to interface with foundries, and when it was first created, GF had no sales team.

GF developed their own 28nm HiK PolySi process, but it was late and the initial market had already gone elsewhere. They skipped 20nm and tried to develop their own 14nm process. They licensed FD-SOI from ST and then, at least to an outside observer, did so little with it that it was unclear if it was just a marketing agreement or a second-source agreement where ST's customers could tick the box that they had an alternative. Eventually, they took that FD-SOI and moved it to 22nm, and even did a JV in China to manufacture it there. They gave up on their own 14nm process, and licensed Samsung's 14nm and did as close to "copy exact" as they could, moving the process from Samsung's Austin fab to GF's Malta fab.

The thing that changed everything was acquiring IBM's semiconductor business. Not so much because IBM was the leader in certain RF technologies that are important in mobile, but because they inherited one of the best technology development teams in the world. GF's 7nm technology (and 12FDX) is completely developed in-house, as will future technologies be.

FD-SOI has turned out to be a niche technology, compared to the FinFET main branch, but it is a very big niche. GF recently announced $2B of design wins in 22FDX. Okay, it's a press release for the sake of a press release on the opening day of SEMICON, taking the design wins they have and adding up the expected production revenues. But it does show that 22FDX is real. In fact, Dan Hutcheson ran a survey two years ago showing a lot of skepticism about FD-SOI (mostly due to lack of an ecosystem around the process) and reran it this year with a more positive outlook. See my post FD-SOI vs FinFET: Dan Hutcheson Re-Runs His Survey.

Gary Patton

So what did Gary tell me? First, he said that now:

the pieces are working as one company. It is night and day from 4 years ago.

They have the dual roadmap in place and feel that they are now on a mission to make it as profitable as possible. With Moore's Law slowing, it is increasingly about differentiation and they have not just FinFET, but FD, RF, and Singapore specialty fabs, so they are well-positioned. Malta (fab 8) is "cranking" on 14/12nm.

Gary echoed what Dan Hutcheson had discovered in his survey:

FD concerns have gone away. There are 47 ecosystem partners, with more to be announced by the end of the year. The ecosystem has switched from us pushing to ecosystem partners pulling.

How about those $2B design wins? Gary told me that 8 have taped out, and more during the rest of the year. There is a lot of noise if you are the first person to tape out, but for any customer picking FD today, there will already be many products in high-volume manufacturing ahead of them. (Gary didn't mention it, but NXP has a whole line of microcontrollers in 22FDX, for example—and they are the leaders in both automotive and security chips).

In the early days of FD-SOI, there was a big push around body bias, and a minor bullet point around RF on SOI. If you need a primer on body bias, see my post Cadence Tool Suite Qualified for 22FDX Reference Flow. Gary told me that the interest has completely switched. You can't do RF on a FinFET process, the gate capacitance (and probably some other things—I am not an RF guy) kills you. But the FD-SOI transistors are much better, and they can build really good power amplifiers. "RF is a cornerstone for FD-SOI". One customer already has 5 RF chips (I'm assuming in design rather than taped out). A big driver is 5G with mmWave.

That $2B number doesn't tell the whole story, we will be back with a higher number.

Those FD-SOI design wins? 7 are cryptocurrency, 16 are IoT, 7 industrial, 3 automotive, 10 networking, 1 mobile. 

GF inherited IBM's RF business too, and they also inherited the mantle of market leader. They have shipped over 35B chips for front-end mobile, over 5B power amplifiers for smartphones and mobile devices (since 2011).

What about 7nm? The first product tapeouts will be in the second half of this year, with production in 2019. There is no EUV in the initial flow. Current pellicles only have 75% transmissivity, and with reflective optics the light path goes through the pellicle twice, so you lose half the power. "You can't do all the work to get to 250W and then lose half of it." So initially GF (and everyone else) will launch with immersion and will insert EUV later. If you don't know what a pellicle is or why 250W is important, then see my post Imec on EUV: Are We There Yet? which is my most recent post on the status of EUV technology.

As I said above, the initial 14nm process was copied from Samsung. The design investment there is huge, so GlobalFoundries are pulling some stuff in from their 7nm development, and doing an optical shrink I'm guessing 10%, which is normal) to 12nm.

Summary

Based on what Gary told me, which of course the market will validate or not over the next couple of years, GlobalFoundries is in a much better place from a technology point of view. They are the clear leaders in FD-SOI, the only company truly committed to it from a manufacturing point of view, with fabs in Germany and (soon) China. They have products in FinFET too, and an in-house-developed 7nm process going into manufacture next year. They are the market leaders in RF with both FD-SOI for fully integrated RF for IoT and 5G, and with the market leading RF business they inherited from IBM.

Of course, they have challenges to win designs against what has to be considered the default choice. But at least it now is a choice. A couple of years ago, Gary told me that he saw his role as two-fold: accelerate the leading edge, and produce differentiated solutions. That is one thing (or two things) that hasn't changed.

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.