SAN JOSE, Calif. — Major changes to design philosophy are need at sub-90-nanometer nodes to accommodate the shrinking number of atoms on each transistor, according to a Monday evening (Nov. 7) panel discussion at the International Conference on Computer Aided Design here.

The six panelists, mostly from academia, proposed radical ideas — including the creation of design tools with artificial intelligence (AI) — for dealing with the less reliable components resulting from the shrinking number of atoms and increase in complexity created by the growing number of transistors.

As the number of atoms in a transistor decreases, effects caused by variation in threshold voltage due to the number and placement of dopants as well as failures created during manufacturing become more pronounced.

Panelists, as well as the moderator, Seth Copen Goldstein of Carnegie Mellon University, agreed that faulty components would be unavoidable and that a successful design will need to account for this.

"We need to build processors that work despite having a large fraction of their transistors in some faulty state at any given time," said Josep Torrellas, a professor in the department of computer science at the University of Illinois at Urbana-Champaign, who suggested that this percentage could be 1-5 percent, or higher.

Tad Hogg, a researcher in the Information Dynamics Lab at Hewlett-Packard Laboratories, said "molecular" electronics would contain many defects and that redundancy would be required to alleviate the problem.

"Instead of being able to ignore the defects and assume you can manufacture perfect devices, you really have to learn to deal with defective devices," Hogg said.

Hogg said the situation created an opportunity for design tools that provide simple abstraction for circuit design in spite of many defects. He called for tools that would predict the probability of producing circuits at a given level of manufacturing technology.

Dan Siewiorek, a professor at Carnegie Mellon, said it is time for the industry to revisit the idea of AI-based CAD tools in response to decreasing time-to-design and increasing complexity. Siewiorek noted that he was involved in the creation of three AI-based CAD tools in the early 1980s and 90s, when the industry faced similar challenges.

Siewiorek said the result was three tools — two created around 1980 and one created around 1990 — that demonstrated results: Weaver, a multi-expert channel router able to switch boxes that were un-routable by conventional approaches; Talib, a rule-based circuit layout synthesizer with more than 3,500 rules; and Micon, a rule-based system using knowledge acquisition to synthesis from specification.

Siewiorek suggested two possible directions for AI in CAD tools, machine learning tools that could recognize similar contexts and apply previous solutions, and data mining tools that could search for design fragments previously used.

"We have lots and lots of design fragments out there," Siewiorek said. "Why can't we mine them, find the thing that is closest to what we need, and start stitching it together?"

Sarita Adve, also a professor in the department of computer science at the University of Illinois at Urbana-Champaign, said designers need integrated solutions that consider system layers and components, failure mechanisms and energy, temperature and performance constraints.

"I believe that we are very good at coming up with solutions for individual pieces of the problem, but as we go ahead we need to do things differently," Adve said, adding that "piecemeal" solutions are unlikely to suffice in sub-90-nm design.

Several panelists emphasized cooperation between architects, designers, EDA vendors and even manufacturing, is key to overcoming these challenges.

"Architects and CAD designers are going to have to be spending much more quality time together, quite possibly at night time," Torrellas said.

Torrellas said designers will need to employ more "architecture-aware" design. Architects, meanwhile, will need to embed redundancy at all levels while abstracting complexity away and without wasting energy, find new ways to quickly identify faulty components and create "self-healing" capabilities for faulty components, all of which will require the help of designers, he said.