Several years ago, there was a lot of publicity about statistical timing analysis, which was thought by some to be the “next big thing” in IC design. Then things got quiet. But as process nodes head into the 45 nm and below territory, statistical static timing analysis (SSTA) may once again become an important new technology.
Rather than focusing on best-case and worst-case corners and reporting absolute numbers, SSTA returns statistical distributions. In an interview I did in 2007 for SCDsource.com, IBM SSTA guru Chandu Visweswariah cited these reasons for using statistical timing:
Multi-corner timing requires too many timing runs for full process coverage.
- SSTA provides full-chip coverage.
- Incremental timing is possible.
- SSTA predicts the sensitivity of circuit performance to all process parameters.
- SSTA offers a yield curve, telling you what percentage of chips will yield at a given frequency.
- Diagnostics for SSTA include yield gradients and critical factors for finding hot spots in designs.
But doubts and questions surfaced about SSTA, some of which were captured in an International Symposium on Physical Design (ISPD) debate I wrote about for EE Times in 2007. Skeptics suggested that “intelligent corner selection” might provide an alternative to statistical timing. Others have said that SSTA is really useful only for random variations, and that modeling is a better approach to systematic variations. Finally, the lack of statistical models from foundries pretty much confined SSTA to IDMs until recently.
Still, EDA vendor support grew, albeit without a lot of fanfare. Cadence brought statistical timing into Encounter in 2007, and today the Encounter Digital Implementation (EDI) system and Encounter Timing System (ETS) provide statistical timing analysis and optimization, along with generation of statistical libraries in the statistical ECSM format. To get a handle on what’s happening with SSTA these days, I talked to Mike Jacobs, senior product marketing manager for ETS, library characterization, and ECSM.
Mike said that adoption of statistical timing is just beginning, and has been slower than some people initially hoped. He noted, however, that “there’s a ton of interest every time we talk about it. People ask a lot of questions. They’re just waiting for someone to come out and say, ‘hey, this is here now.’”
One reason it might be “here” now is that foundries are beginning to support statistical models. TSMC, for instance, supports SSTA in its reference flow 8.0. Another is the growing impact of process variations at 45 nm and below. As Paul McLellan noted in a recent EDA Graffiti blog about statistical timing and process variations, “manufacturing is a statistical process and isn’t pass/fail at all.”
While there’s some interest in SSTA at 65 nm, Mike said, most people are looking to 32 nm as the node at which statistical timing will become mandatory. At 45 nm, according to analyst Gary Smith, it’s being used more selectively. “Instead of coming into widespread use at 45 nm, as we expected, it [SSTA] is being used in just the most critical applications,” Gary said. “Just as we ran Spice on our critical paths fifteen years ago, we are using statistical timing on our critical blocks today.”
One major use for SSTA is determining how well a design will yield in silicon. It may show that a given design is “good enough,” saving an expensive rework effort. Mike related how one customer had an ECO that affected the netlist, and was concerned about the impact on negative slack. By running SSTA, the customer was able to determine that the impact would not result in unacceptable yields.
What about the learning curve? “If you can get your head wrapped around the probability of a violation versus a discrete negative slack value, that’s all you really need to do,” Mike said. “Running the tool is really very simple. It’s just like static timing, except the reports are a little different.”
The real challenge behind SSTA is the paradigm shift it involves. IC designers, like quantum physicists, must learn to think in terms of probabilities. Albert Einstein reportedly said, “God does not play dice.” Perhaps – but deep submicron silicon on a foundry assembly line does.