Post a Comment
News and New Products
Characterization tool aids SSTA-library creation
By Michael Santarini, Senior Editor -- EDN, 1/15/2007
To account for process variation while eking out the best mix of performance, power, and yield from new digital-IC-design processes, the IC-design industry is now moving from STA (static-timing-analysis) tools to SSTA (statistical-STA) tools. After years of researching SSTA, companies such as IBM, Cadence, Synopsys, Exteme DA, and Magma Design Automation now offer SSTA tools commercially. But to make those tools work properly requires new timing libraries that account for process variability, not just worst-case-timing estimates. Toward that end, EDA start-up Altos Design has released a library-characterization tool to help foundries, IDMs (integrated-device manufacturers), library vendors, and fabless-IC vendors create their own SSTA libraries for several commercial SSTA tools.
Altos Design introduced its first commercial offering in July 2006. That tool, Liberate, is for characterizing I/O and standard cells for static-timing libraries and claims runtime improvements over current commercial offerings. Now, the company is introducing Variety, a variation-aware SSTA-library-characterization/modeling tool that can generates libraries for various commercial offerings.
The vendors offering SSTA tools also offer characterization tools, but each one generates libraries only for its proprietary format. For example, Synopsys’ PrimeTime XT SSTA tool uses a new version of Synopsys’ CCS (Composite Current Source) format, and Cadence’s statistical tool uses Cadence’s statistical S-ECSM (sensitivity-based effective-current-source-model) format. Similarly IBM, Magma, and Extreme DA have their own formats.
Jim McCanny, Altos’ chief executive officer, says that Variety, as its name implies, will generate libraries for many of those vendors’ formats and will do so faster than each vendor’s characterization technology. “With Variety, users will be able to generate libraries that account for both systematic and random variation,” says McCanny. “Other tools take a black-box approach. We don’t do that. We try to understand the circuit. We understand the functions, what vectors are required to sensitize, what paths there are; we do a transistor-level preanalysis of the circuit.”
Whereas models for STA tools capture worst-case cell performance, Variety SSTA models characterize each transistor in a cell and account for both systematic and random variation. The tool generates SSTA models with nominal timing information plus additional data representing the impact of any number of parameter variations. McCanny notes that systematic variation is relatively easy to characterize, whereas random variation tends to be more complex. He says that a typical systematic variation is a shorter gate or trace length. “If, for example, a length shrinks by 5%, we can characterize it and calculate the sensitivity of the delay to that change of length,” says McCanny. When users add a systematic variation to the characterization, it typically adds one times the runtime to the tool.
Characterizing random variation is a bit more challenging. “With random-variation characterization, we model the effect of a parameter on each unique transistor in a cell,” he says. “An average standard cell has 25 transistors per cell, so if you did this with a black-box method, you would end up requiring about 25 times more characterization effort. That’s impractical, so what we do is to reduce that 25 times to somewhere around two to three times.”
The tool employs an algorithm that quickly locates the transistors that a user-specified random variation will affect. “You can model any Spice parameter you have in your model: low- or even high-level parameters you created from a combination of physical parameters,” says McCanny. Users can add any number of variations to the characterization runs. He notes however that, each time users add a random variation parameter to the tool, it adds three times to the runtime.
Renesas has benchmarked the tool modeling length, width, and thermal oxide as systematic variations and threshold voltage as a random parameter. It characterized 387 cells in six hours on an eight-CPU system. It characterized nine systematic variations and one random variation in 27 hours on a 16-CPU system.
Variety creates libraries in its own .lbd database format and then translates the file into ECSM, CCS, or Extreme DA’s internal format, with other formats to follow. That translation allows users to use the same library with any of these vendors’ tools. McCanny notes that it will be especially useful for design groups who use one vendor’s SSTA tool for design and then use another vendor’s SSTA tool for sign-off. Characterizing libraries in Variety allows both tools to work from the same library.
This neutrality has led Altos to offer Variety to the Si2 (Silicon Integration Initiative) to help the organization’s OMC (Open Modeling Coalition) derive an industry-standard reference flow for library characterization, modeling, and model usage. The Si2 will install Variety into OMC’s reference flow as the characterization subsystem for statistical characterization of library data. As part of the reference flow, Variety and Altos will help the OMC define standards for library-flow interfaces for communicating statistical information between elements in the flow.
Variety’s entry price is $95,000, and prices for a typical system start at approximately $250,000 for a one-year subscription. The tool currently supports Cadence’s and Extreme DA’s SSTA tools, and Synopsys’ tools will soon follow. The company hopes to latter add support for Magma’s and IBM’s SSTA tools.
