Chandu Visweswariah, along with Kerim Kalafala, lead a team of R&D engineers at IBM Research and IBM Electronic Design Automation. Their small group recently won EDN's Innovator of the Year award for developing the EinsTimer statistical-timing tool, which also took top honors in the EDA Design and Implementation Tools category. Their group also won the Design Automation Conference's best-technical-paper award in 2004, and the company commercially released the product before the Design Automation Conference in 2005.

What is statistical-timing analysis, and what problems does it tackle?

Statistical-timing analysis traditionally handles process variations in a corner-based manner. In other words, the timer checks chip performance and timing relationships at various discrete "process corners." The main problems with this approach are that, in modern technologies, the number of corners can be large, and the technique does not lend itself to robust optimization. Statistical-timing analysis models gate and wire delays as probability distributions with complex correlations and predicts chip performance and parametric yield as probability distributions. The advantages of the new paradigm include fast turnaround, incremental operation for optimization, pessimism reduction, sensitivity prediction, and enablement of performance-versus-yield trade-offs.

Analog designers are used to statistical proofs, but why does the digital world need them?

Digital designers relied on being able to bound their performance in a relatively straightforward way. But, now that wire delays play a larger role in determining performance and metal levels exhibit independent variations, it is no longer easy to bound performance. Because variability is a complex, multidimensional phenomenon, solid tool support is necessary to help circuit designers to maximize both parametric yield and performance.

Why will the technology become important in the future?

As you move to 65-, 45-, and 32-nm processes, variability will only get worse. As you reach fundamental physical limits, you'll see the statistical nature of matter in dopant fluctuations, oxide-grain boundaries, and line-edge roughness. Making individual worst-case scenarios of all these sources of variation will induce pessimism that will rob the new technologies of any performance gain. Besides, robust design and adaptive techniques will become more common. Hence, it is imperative that you embrace a statistical design-and-optimization methodology.

You said that launching the research program was difficult and required buy-in from management. What does management now think?

Management is enthusiastic! Any new paradigm generates resistance and inertia, particularly if a delay or arrival time is no longer a number but a probability distribution instead! However, with the right combination of generating results, persistence, and perseverance, statistical techniques are finding their way into mainstream use, including in IBM's ASIC tool kit. IBM is introducing these techniques in a phased manner, and they are invaluable parts of accurate timing with reduced pessimism and fast turnaround. Better process modeling couples with statistical timing to be the best way to achieve realistic timing results.

When IBM last year announced the tool, the company said it would offer it commercially to all comers and that it would be fab-independent. Is that still the case?

IBM has enabled EinsTimer statistical timing in a fab-independent manner and has been working with a few clients in this mode. However, IBM's main focus is on providing comprehensive approaches to its technology and platform partners. Regardless of where customers fabricate their chips, IBM offers its clients design services that can take advantage of the company's knowledge and expertise, including statistical timing and optimization.

What are the next steps for the technology? How do you proliferate the technology and get designers not only up to speed on its benefits, but also able to use it on real designs?

Several parts of the methodology must evolve to take full advantage of the new statistical paradigm: process modeling, library characterization, physical synthesis, and test. Several factors contribute to quick proliferation. First, the tools must be pushbutton and easy-to-use. Second, the timing reports should be intuitive and configurable so that someone schooled in deterministic timing will easily get used to the new reports. Finally, training, education, support, and documentation will play big roles in widespread acceptance of the new tools.