So far the move to lower semiconductor process nodes has continued unabated, but the upcoming 20nm node is causing a lot of concern. Lithography is so challenging that extra masks (double patterning) will be required. Will designs be technically and economically feasible? Panelists at the Design Automation Conference (DAC) June 6 said "yes," but acknowledged that the challenges are very real.
A Cadence-sponsored panel at DAC, titled "Getting a Jumpstart on 20nm," explored the challenges of this emerging process node in depth. Panelists (shown left to right below) included:
Setting the tone for the discussion, Handy commented that "20nm is really a very different process node. Some say it's shaking the very core of semiconductor design. You're going to see new processes and techniques." Following are brief answers to some of the questions posed to panelists.
Q: What's the single most difficult thing about 20nm?
Segars: "What bothers me most is pure economics. With double patterning, masks, and NRE costs, will it make economic sense to design at 20nm?"
Hunter: "The economics are challenging, but customers are asking for it sooner, not later. There's still a lot of interest in getting more performance out of 20nm. The challenge is bringing everything together fast enough. A lot of development work on the process side and the EDA side needs to be done."
Magarshack: "We could spend five years optimizing libraries and design rules, but we have to get going. I think the main challenge is to find a way to stop negotiating design rules."
Hsu: On the business side, "EDA is hard-core software development that takes a lot of resources." On the technical side, turn-around time will be the challenge. "We will figure out how to make use of the process. The question is how well we can take advantage of the process."
Q: At 20nm we're doing high-k metal gates (HKMG) and low-k metal interconnect. What do we need to do to avoid problems?
Segars: "We're trying to abstract out all that horror of high-k, low-k, so designers only need to worry about the architecture and the implementation."
Magarshack: High-K, low-k combination is very positive. High-k transistors gives performance boost of at least 20 percent. Low-k wires lower capacitance, providing lower power or higher performance.
Hsu: "As long as you can make the process work, and the libraries work, we will make [EDA] work." But early collaboration is crucial. Cadence started working on 20nm much earlier than with other process nodes - the effort started 2 ½ years ago, in fact.
Q: Double patterning gives you two times as many transistors per a given lithography, but it needs highly regular structures. What are the challenges in using double patterning?
Segars: "It certainly imposes new design rule constraints. You have to make sure shapes can be split into two masks. As library guys, we try to abstract away those details."
Magarshack: "We need to collaborate early with the library and process guys to get a better compromise than a simple, pushbutton solution that would leave something on the table in terms of density and power."
Hsu: "All of our partners use slightly different approaches. Some use double patterning, some don't, some add different variations." Also, double patterning has an impact on electrical characteristics and variation that must be considered.
Q: Is double patterning necessary at 20nm?
Hunter: Samsung has chosen to optimize its process for double patterning to get better density. Going below an 80nm pitch requires double patterning. We have chosen to go with a 64nm pitch. This will provide about a 35 percent performance improvement from 28nm to 20nm at the same leakage.
(There was no specific question about 3D-ICs, but several perspectives were given).
Hunter: 3D-IC is very complementary to 20nm and we are investing in it heavily as well. High voltage or analog circuits may not need the shrink path of digital.
Segars: "We've tried to do every type of circuit by putting all transistors on one die. If we can split up the transistors into 65nm analog/mixed-signal and optimize digital transistors differently, this [3D-IC] technology could really help."
Magarshack: "So far we have not found an economically viable product for TSV [through-silicon via]. We're still at LPDDR2 and LPDDR3 rather than jumping into wide I/O."
Having listened to the panel, I have no doubt that foundries will offer a 20nm capability and some people will use it. But how many and how soon, and how well will libraries and tools take advantage of the process? Very early and deep collaborations, such as those undergone by Cadence and the other companies represented in this panel, are the key to 20nm success.
Kudos to Richard for accurately capturing the important points of the panel.