I went to a briefing last Tuesday where Samsung and GlobalFoundries announced that they have the same process at 14nm. Dan already wrote about itso it is old news in one sense. But I really think people underestimate its importance. In essence, reading between the lines, Samsung is licensing GF their 14nm process. This is driven by large customers (of course officially nobody can say who but Qualcomm and Apple would be good guesses) who don’t want all their eggs in the Samsung basket, especially since Samsung competes with them in the biggest end-market of them all, mobile.
How similar are the processes? They just have one PDK for both companies, the designs are completely portable. I don’t know whether GF is doing a sort of Intel copyexact whereby they put everything into their Malta NY fab just the way Samsung has in the Austin TX fab. The presentation says that all fabs are in sync involving materials, process recipes, integration and tools.
I think this is really big news. GF has only recently started to really ramp 28nm which means that TSMC had a free run for a couple of years and was really the only game in town for most people (yes, Apple manufactured at Samsung but the average customer isn’t going to get Sammy’s attention). This meant that there was no price competition at 28nm plus a lot of systemic risk when only one company is building so much (if one of their fabs was damaged in an earthquake or a fire, for example).
At 14nm GF/Sam there will be 4 fabs that can build 14nm wafers. Samsung’s S1 in Korea for prototypes, and then S2 (Austin) and S3 (Korea) and GF’s fab 8 in Malta New York. 28nm will continue to be run mainly in Dresden Germany. I didn’t think to ask whether Fab 8 has actually run 14nm silicon yet although they did have a 14nm wafer at the presentation (although to be honest, a wafer is a wafer, it’s not like FinFET looks different to the naked eye).
Another important fact is that Samsung/GF claim that their area is more than 15% smaller than “other foundries” for which we can read TSMC. The BEOL (metal fabric) on the process is the same as 20nm so I don’t know if this 15% is very dependent on how much memory is on the chip. They say they have the smallest memory solution and innovative layout schemes for compact logic.
I asked if this was a cost-reduction node and nobody would commit that it was. Clearly the process has lots of value for some markets with lower power, enormous density and so on. My rule of thumb for process technologies, though, is that if nobody says the cost per transistor is going down then it isn’t (or at least not much).
What is the timing on all of this? 14LPE is already qualified. 14LPE and 14LPP (more performance, lower power) PDKs are already released. MPW shuttles are available (presumably in S1). Prototyping now. Volume production by end of the year.
GF had process development going on already for 14nm and that is planned to continue. I’m not sure that will turn out to be true, if all the customers opt for the two 14nm processes covered by this announcement. GF has had almost unlimited amounts of money from Abu Dhabi but just money to spend doesn’t get you a competitive process in a competitive timeframe.
The reason that this is such a big deal is that, based on their performance at 28nm, GlobalFoundries was uncompetitive, too late to market to get more than the crumbs from TSMC’s table. Any semiconductor company only makes money if its fabs are close to full and theirs was not. If they didn’t get competitive by 14nm then the future of the company was dubious. Abu Dhubious.
This deal with Samsung means that they are back in the game. At 28nm, TSMC had over 80% market share. At 14nm (TSMC calls it 16, process names are just silly these days since there is nothing 14 or 16nm around) it will be more like 50% since I expect that the high volume customers will spread their love around.
More articles by Paul McLellan…
AI Semiconductor Market