Experts At The Table: Multipatterning

Last of three parts: Manufacturing needs vs. EDA’s risk appetite; limited value of the ITRS roadmap; using colors effectively in double patterning; apportioning blame if something goes wrong.

popularity

By Ed Sperling
Semiconductor Manufacturing & Design sat down with Michael White, physical verification product line manager at Mentor Graphics; Luigi Capodieci, R&D fellow at GlobalFoundries; Lars Liebmann, IBM distinguished engineer; Rob Aitken, ARM fellow; Jean-Pierre Geronimi, CAD director at STMicroelectronics; and Kuang-Kuo Lin, director of foundry design enablement at Samsung Electronics. What follows are excerpts of that conversation.

SMD: Are the tools that are available today sufficient to deal with the challenges at hand in terms of manufacturing and packaging?
Geronimi: It’s time to move from a custom solution to an engineered solution.
Aitken: If you look at double patterning, there’s a routing challenge, especially on metal-one. You go through so much effort on standard cell design and you design something that’s decomposable using the various coloring rules that there’s no way you’re going to let the router touch it. Routers don’t understand enough about design rules to be allowed anywhere near it. That is an inherent inefficiency. You lose a fair amount of density by not letting routers onto metal-one. There are similar tool gaps for what metal corners look like. There’s another one in terms of electromigration and IR drop. There are a lot of tool gaps and designers are working around them, but the inefficiencies are there. Better tools could help eliminate some of those inefficiencies.
Capodieci: With respect to multiple levels of patterning, although we’re doing a lot of joint work there I haven’t seen a robust solution being deployed. We need to attack the problem of multi-patterning to realize that problem will be solved by solving another problem—from decomposing things into manufacturable blocks. For directed self assembly or direct-write, some other form of non-patterning approach will be used. Multiple patterning is a useful exercise. I don’t think it’s going to lead us toward an industrial solution because triple and quadruple exposures and all the processes associated with that have technical and financial problems.

SMD: The design industry looks at what can be created, while the manufacturing side looks at it from the standpoint of whether it can be produced with reasonable yield. As business considerations begin creeping into both sides, how does that affect this relationship and the decisions being made?
Liebmann: There’s an increasing reluctance, at least from the tooling side, to bite off more than is digestible. We’ve talked about construct-based design. That extends into construct-based routing. I haven’t found any takers for a router that can deal with 100 legal configurations and nothing else. That’s a very long and expensive development project that the EDA industry won’t take. So we’ve had good luck making improvements to the existing design base. But to really make revolutionary changes on the design side is just as hard as making evolutionary changes on the manufacturing side.
Geronimi: There are things we need to consider now. If we don’t get the tools to work we will have much more difficulty later on.
Liebmann: A lot of this comes back to the lack of a clear roadmap. The ITRS roadmap has become something of a useless document at this point because it extrapolates to zero without taking into account these very disruptive things that are happening, such as double patterning, maybe the switch to EUV, and the introduction of self-assembly. Without a clear road map it’s difficult to ask the EDA industry to invest in a six-year project to develop a new routing technology.
White: It is challenging, with finite resources, to invest a huge chunk in a product development team that can go off and do something that will take multiple years to create a solution. What you find within the industry is that it’s easier or more straightforward to improve on the infrastructure that we know and extend it. That’s the general strategy you’ve seen from all of us.
Lin: Right now most of the burden of double patterning is on the foundry and EDA. I would like to encourage the fabless design houses to look at the different options, not just to rely on EDA and the foundries.
Aitken: What we found with double patterning is that there are two kinds of foundries in the world. There are those that want to see colors and separation in the layout and those that do not. We find it’s actually harder to do the ones that don’t separate because what we do is add the colors, to prove that it can be separated, and then we erase the colors and hope they can find the same decomposition that we did. I’m not convinced that’s the optimal way to do something. Decomposition is a major challenge. Double patterning is at least solvable in normal terms. Triple patterning will require heuristics.
Liebmann: Sometimes marketing gets in the way. There is a fear of inconveniencing the customer, which in this case is the fabless design house. There is no way around it. If you do double patterning, the designers have to get engaged and understand the fundamental problem. This whole idea of coloring is just one way where you need to convince the customer there is a problem, but once they understand that I think it’s a very solvable problem.
White: As we work with customers doing these kinds of designs we are seeing a convergence. Industry-wide, there’s a growing tolerance of coloring. In many instances it’s important information that otherwise is being lost.
Capodieci: We cannot be color-blind. We must see in color.

SMD: Who’s responsible as things go wrong with a more complicated design?
Capodieci: The collaborative model that many of us have been putting in place even at 28nm, where there is very limited usage of double patterning, will be the only way forward. Part of the color awareness is that. It’s not just exchanging operational information about the coloring. It’s also understanding what the corner cases are with the IP and with the routing. Rather than taking the responsibility for blame, we all have a shared responsibility for success. That also means we have a shared responsibility for failure.
Lin: It helps that we have more pieces internally and are working with our external customers. Our early learnings will help us to work with customers earlier.
Capodieci: The golden rule is, ‘When in doubt, blame the vendor.’ If you are an IDM, you blame whoever is external. If you are in a more complex ecosystem, whoever is holding the vendor badge that day gets the blame. Triple patterning brings to the surface a very complex interaction. There are EDA, the foundry, the design customers, and a very stratified group of IP vendors. So maybe there is not a single solution.
Aitken: It’s a very important question. When you look at it historically, we’ve evolved a number of different handoff points. The DRC deck is a classic one. It’s either legal or it’s not. If it fails and it was legal, it’s the foundry’s problem. If it fails and it was illegal, it’s your problem as a layout person. There’s growing recognition that’s not adequate. We want it to be DRC-legal, but we also want sufficient yields, so we check and make sure the things we think are DRC-legal will yield even if the DRC says they will. That level of collaboration will be necessary going forward. The business arrangements haven’t quite figured out how to track it yet, but they’re going to have to. The ‘blame the vendor’ approach is one way to do it, but if it’s your part that isn’t shipping then you’re the one getting fired and that doesn’t matter. So it’s everyone’s responsibility.
Capodieci: That’s exactly right, and that’s why we need to talk about shared responsibility. Regarding the example about DRC and additional yield assurances, many problems today make DFM verification decks mandatory. You must achieve a certain score. There is an open pass/fail, based on the fact that you’d like to have an 85% score on your recommended rules. How you achieve that score depends on which geometries you want to push and which ones you want to relax. This can be seen as an additional level of complexity, but it’s also an additional level of freedom. You can play around with your physical design and still get to tapeout.
Lin: You have to define a very thick interface and define things very clearly. Within the same company you can resolve the problem over lunch. Here you need a very clear explanation.
Geronimi: Unfortunately with double patterning, if you cannot get to signoff you cannot manufacture it.
Aitken: I disagree. When you have some very complicated structure you can sign off and say it passes, and you have no real assurances that it works. That’s where the challenge is.
Geronomi: From a double patterning standpoint, when you look at whether you can decompose it or not, you go to manufacturing with colors expecting the rules to work correctly. At least with single patterning you know that even with very complex rules, you can manufacture the design.
Liebmann: That, to a large extent, is why double patterning is in better shape than DFM. DFM is very vague. Who owns it? Is it really required? Is it optional? With double patterning, it comes down to a double patterning-enhanced rule set. You either pass DRC or not. It’s very clear.
Aitken: but there are some challenges there, too. If you create decomposition and send it to a fab, suppose they recolor it and it fails. Who’s problem is it?
Capodieci: That’s why we have been expanding the safety net of DFM to include additional checks for double patterning. We had to add scores for double patterning. Plus, we are finding that a lot of patterns are derived from decomposition. DFM will have an enhanced role to catch everything to catch the holes left by double patterning. At the end of the day, double patterning needs to come together to create a single mask for future processing. That’s the key. It’s easy to break. The question is, when you put it back together will it yield?