After reviewing the myriad issues that have brought DfM to the forefront in EDA tool planning, Kahng observed. "There is a fixed trajectory in which our industry deals with problems. First, we attempt to make the problem go away with new design rules. That quickly proves insufficient. Then we develop detailed models of the problem mechanism. That quickly proves to be overkill. Finally, we incorporate the solution to the problem into our libraries."

Kahng and Gupta in effect suggested short-circuiting the trajectory in the case of DfM by taking solutions directly into design libraries. They illustrated the approach with potential solutions to several major DfM issues.

First, they addressed the issue of exploding file sizes in sub-130 nm mask data sets. Kahn said much of this complexity could be reduced simply by adding only optical proximity correction features, sub-resolution assist features and phase shifters to features where they are necessary to ensure circuit performance, not to every feature.

For example, he said automatic tools that add hundreds of features to guarantee that a rectangle of dummy metal remains rectangular contributes nothing the behavior of the chip. But at a more detailed level, Kahng said it was possible even with existing databases to attach to each feature a number conveying the accuracy the designer requires — thereby cluing reticle enhancement tools about how much effort to expend on perfect reproduction of the feature.

"We should work in the direction of determining how aggressively to correct a feature's shape based on its impact on the standard deviation of delay," he said.

Gupta extended this notion to the concept of applying optical proximity correction (OPC) to library cells when they are created, rather than attempting to apply OPC to the entire chip after placement. He said the radius of a feature's influence in 193-nm lithography is no more than 600 nm, but typical library cells have 200 to 300 nm of empty space at their periphery.

If cells were designed assuming some "typical" environment at their borders, corrected for sub-wavelength optical effects individually and then placed without further processing, the only effect should be a slight distortion of features lying at the edges of the cells. This, he continuedm, is unlikely to have a major impact on circuit performance.

In fact, Gupta said a subset of a cell library had been created and tested in this manner and had yielded promising results for preserving critical dimensions. "There was some difficulty with contact coverage at the edges of the cells," Gupta said, "but we are examining this."

Kahng illustrated the concept of library-based solutions by examining the problem of phase conflicts in minimum-dimension gate lithography. The problem involves resolving minimum-dimension gates, hence it is necessary to place a phase shifter on each side of the gate feature on the mask set, one shifter in each polarity. When a series of gates are close together, it may be impossible to arrange the shifter rectangles so they don't overlap and conflict with each other.

Kahng said designers have two alternatives to remove phase conflicts: move one of the gates out of the way so that the phase-shift rectangles no longer interfere, or make one gate bigger to eliminate the need for phase shifting. A tool to perform this analysis on an already-placed chip would face an imposing task, and would need clear data on design intent in order to work.

Instead, Kahng proposed building cell libraries in a way that would produce phase-correct-by-construction placements. Cells would be designed with the phase shifters — like OPC features — already in place. Cells with shifters near their boundaries would have varied designs so as to allow picking a cell of a given function with a given arrangement of phase shifters at its boundary.

The concepts fit together into a clear message: DfM is now so complex that post-processing of an entire mask set is impractical. Instead, solutions have to be incorporated into cell libraries, and applied during placement and route by tools that are aware of the designer's intent. "No one of our diverse communities — designers, EDA developers or foundries — can solve this alone," Kahng warned. "We will have to get their through coevolution."