By Ed Korczynski and Pete Singer, SemiMD and Solid State Technology
Fabless companies could skip the 20nm node and move straight to 14nm FinFETs. That is the hope of GLOBALFOUNDRIES and Samsung who are announcing a joint program that offers a single process design kit (PDK) and manufacturing at four different fabs with identical processes.The PDKs are available now, and 14nm manufacturing could move into high volume production by the end of the year.
“This is unprecedented,” said Kelvin Low, senior director of marketing at Samsung. “It never has happened in the industry, especially at the very leading edge nodes. We are confident that this will transform the supply chain model,” he added.
Fabless companies such as Qualcomm have been lobbying for such multi-sourcing for some time, and are eager to move to FinFETs which offer higher performance and reduced power consumption. The 14nm FinFET offer a 20% improvement in performance and a 35% reduction in power compared to 20nm technology.
“For both Samsung and GLOBALFOUNDRIES, we will be providing our customers with a choice and assurances of supply, enabled by the unprecedented global capacity across our respective manufacturing facilities throughout different locations worldwide,” Low said. “For Samsung, we have facilities in the U.S. in Austin. We also have a couple of plants in Korea. For GLOBALFOUNDRIES, the 14 nm capacity will be in Malta, NY.”
The single process design kit will allow customers to do a single design that is capable of being multi-sourced from different locations.
“This really is a change from the existing supply chain model where customers are trying to design multiple designs to multi-source their products,” Low said. “True design compatibility in this collaboration will allow customers to better manage their design NRE and they can focus on bringing the product to market on time. Both companies see this as a necessary advancement of the supply chain model and we start off with the 14nm LPE as well as the 14nm LPP technology node.” 14nm LPP is a follow-on offering which has additional performance enhancements as well as power reduction attributes.
Samsung had already developed much of the process technology for LPP and LPE flows to run using 14nm node finFETs, while GLOBALFOUNDRIES was working independently on another 14nm node process variant. The two companies decided to pool resources to save both time and money in bringing 14nm node finFET capability to the commercial IC foundry market.
“Because of customer interest in having that assurance of supply and being able to do one GDS and being able to work off of one common PDK and source at both of our companies, we decided to work together and go with the 14 LPE and 14 LPP as common offerings between Samsung and GLOBALFOUNDRIES,” said Ana Hunter, VP of product management at GLOBALFOUNDRIES.
Low said that Samsung is already running 14nm products in its fab in Korea. The 14nm LPE, for example, was qualified earlier this year.
“We are already in silicon validation of our lead customer products. We expect to ramp production by the end of this year,” he said. Design activities started almost three years ago. “Right now, we are seeing a lot more pickup overall by the lead adopters and even other customers following suit, mainly because the marketplace does see that the 14nm FinFET is at the right maturity level for volume production,” Low said.
Although there is still lot of activity at 28nm, the technology is considered to be in a mature phase. “We still continue to see healthy, new design-ins,” Low said. “Of course, there are a lot of requests to see what additional enhancements we can do at our 28nm node to prolong the lifespan of that node.”
What about 20nm? “From Samsung’s viewpoint, we see that a relatively short-lived node, mainly because of the overall resonance of FinFETs and the eagerness of customers to migrate from 28nm directly to 14nm FinFETs.”
Hunter agrees, noting that 28nm has been in high volume production for several years now. She said GLOBALFOUNDRIES does have 20nm product tapeouts running in the line, but said that she does not see 20nm being a very extensive node in that most customers are eager to get onto FinFETs.
“We do have products running at 20nm, but I think the design efforts will quickly go over to FinFET and we’ll see that be a much longer lived node with a lot more product tapeouts,” she said.
The companies say the 14nm FinFET offering could be up to 15% smaller than that available from other foundries due to aggressive gate pitch, smaller memory solution and innovative layout schemes for compact logic.
Hunter, having been a VP at Samsung before holding her current position at GLOBALFOUNDRIES, noted that the two companies, along with IBM, have been in collaboration for quite some time on “The Common Platform” at 65, 45, and 28nm nodes, but this announcement is strictly between GF and Samsung.
“We do continue to work with IBM in other areas at the Albany Nanotech center, where there is continuing collaboration on more advanced nodes, on materials research, pathfinding, and advanced module development kind of work,” she said.
Fabless customers use a single PDK to do a single design, allowing a single GDS file to be sent to either company. The design-for-manufacturing (DFM) and reticle-enhancement technologies (RET) needed at the 14nm node are challenging.
“We go deep into the collaboration, even to the OPC level and a lot of sharing on DFM as well. It is a very extensive collaboration,” confirmed Hunter. “At 14nm the designs are extremely complex, and to be able to truly supply multi-sourcing from one GDS, you have to have that level of collaboration to ensure that the output from all of our factories is the same. That’s a huge advantage to customers because the idea that you could source from two different companies without the kind of collaboration that GLOBALFOUNDRIES and Samsung are doing is just simply impossible when you get into 14nm FinFETs. When you get into the complexity of the designs, the databases, the amount of reticle enhancement techniques that are required to be able to print these geometries, you need to have that kind of in-depth collaboration.”
Low said that the two companies have a “fabsync” structure running in the background to ensure the fabs are fully synchronized.
“There are a couple of things we are doing proactively,” he said. “The technology teams are deeply engaged with each other. We have technology workshops across both companies. We have test chips that are run regularly to ensure that the process continues to stay synchronized. These test chips are not just simple transistors. We have product level elements that we’ve included to make sure we measure the critical parameters. This is only enabled through open sharing of technology information.”
Hunter adds: “We run the same test chips, we share wafers back and forth to measure each other’s products to make sure all of our equipment is calibrated, test equipment calibrated, results are the same on exactly the same test chip.We have test-chips with product-level structures that run in all fabs and both companies share all data to ensure that all fabs stay in alignment. Not just SPICE models and SRAMs, but full chip-like design features.”
However, customers will have to re-do lithography masks if they want to move manufacturing from one company to the other, in part because of issues with shipping masks. Kevin Low, Samsung’s senior director of marketing, commented, “We’ll be providing our customers choice and secure supply. At Samsung we have capacity in Texas and Korea.”
Cost/transistor for 14nm may not be lower compared to 20nm and 28nm. Hunter said, “To continue with optical lithography, it is challenging to do double-patterning and keep costs low.” However, since much of the motivation in moving from 20nm to 14nm is for power-sipping mobile SoCs, by reducing the power consumption by the claimed 35% there could be cost-savings at the packaging level such that the overall product cost is reduced.
To be able to offer essentially the same manufacturing process to customers, GLOBALFOUNDRIES and Samsung had to harmonize not just process recipes but many of the OEM tools used in these fabs.
Hunter says, “To get the same results at this node, it does require engineering down to the tool level and the individual recipe level. That doesn’t mean all tools are exactly the same, however, since cost and availability of tools may have been different when the fabs were equipped.”
Customers can choose which foundry that choose to work with, and then they can choose to discuss commercial terms such as which specific foundry site may be booked to do the work.