IBM claims edge on EDA guard in statistical tools

Santa Cruz, Calif. — IBM Corp. is making its most ambitious foray to date into the electronic design automation market with EinsTimer, a commercial spin of the statistical timing tool that the company has been using internally for roughly a year. The suite is available for licensing from IBM's Engineering and Technology Services (ET&S) division, which will offer it along with chip design services.

IBM has been down the commercial EDA road before, with limited success. The company hopes for a better result with EinsTimer, which it calls the first commercially available tool for incremental statistical timing analysis. The tool can be easily integrated into existing design flows and does not require the use of IBM fabrication facilities, though they are available to EinsTimer users, IBM said.

In an era of variability, statistical timing promises to maximize chip performance and yield. Instead of single numbers, it returns probability distributions, sensitivity graphs and yield curves. It can tell a designer how well a design will yield across a frequency range or predict worst-case performance across process variations.

However, there is disagreement over who needs statistical timing and when. And the announcement puts IBM into competition with EDA vendors that have already announced plans to launch statistical timers, including Magma Design Automation and startup Extreme DA Corp., and may set the stage for future competition with Cadence Design Systems and Synopsys.

But IBM is confident EinsTimer will prove compelling. "This is really the industry's first incremental statistical timing analysis capability, including support for both front- and back-end environmental variations," said Dale Hoffman, CTO of IBM Engineering and Technology Services.

EinsTimer has been IBM's gate-level static timer for more than a decade. The tool's statistical capability has been in use at IBM for about a year, and the company has been phasing statistical methods into design kits over the past three years.

Can IBM translate that into commercial EDA success? The company in the past has fielded such EDA technologies as timing analysis, test and, most recently, formal verification, but it has never amassed much market share in the design tool market.

"If IBM really wants to succeed as an EDA vendor, they can," said Gary Smith, chief EDA analyst at Gartner Dataquest. But the industry giant's past attempts at commercializing EDA technology have been "half-hearted," he said.

Smith called EinsTimer's statistical capability "state of the art" and said statistical timing will be needed by IC layout teams at 45 and 65 nanometers. "They picked a good tool this time," he said. The open questions are whether it can be used by non-IBM engineers and whether IBM will adequately market and support it.

The latter is the domain of IBM's ET&S division, which was founded in 2002 and today employs 1,300 engineers. "A software-licensing solution is complementary to our services offering," said Hoffman. But "we don't think users want point tools; they want solutions," he said. "Software licensing is just one component of what we offer."

The EinsTimer suite includes gate- and transistor-level static timing analysis, statistical timing analysis, power distribution analysis and signal integrity. It comes with a characterization tool that can work with data from commercial foundries.

Startup Extreme DA claims to be close on IBM's heels. CEO Mustafa Celik said his company's statistical timing tool has reached production quality and will probably be announced in July. The offering will include a statistical extractor and a library characterization tool, with signal integrity soon to follow. Celik said that his seven-person startup is not worried about competition from EinsTimer, which he views more as a "point tool" used inside IBM than as a full statistical environment.

Magma's Quartz SSTA statistical timing analyzer, which includes extraction and library characterization, is in alpha release with early customers, and production release is expected later in the year, said Suk Lee, general manager of Magma's silicon signoff business unit. "IBM cannot offer an integrated, competitive solution," Lee said. "Because it's targeted for IBM's flow, IBM will have difficulty commercializing EinsTimer."

Meanwhile, there are questions about how soon statistical timing will be needed. While layout engineers need it at 65 nm, design engineers won't need it until 45 nm or below, said Dataquest's Smith.

"We expect to see statistical timing methods appearing in 2006 for very high-end-performance designs at 65 nm," said Eric Filseth, vice president of marketing for digital implementation at Cadence Design Systems Inc. "A huge amount of silicon infrastructure and supply chain work needs to be done in the meantime."

Statistical timing analysis will initially target 65 nm and see mainstream use at 45 nm, said Rajiv Maheshwary, senior director of marketing for signoff and power products at Synopsys Inc. He said Synopsys is "partnering closely with customers" on variation-aware analysis and extraction.

Statistical analysis is of interest because nanometer IC designs are subject to high process variability. Designers can control variability by applying guardbands, but they will give up performance to do so, statistical timing advocates say. Statistical tools can examine all the possible corner conditions created by process variations.

"Statistical timing is something we've needed for quite a while," said Chandu Visweswariah, manager of circuit and interconnect analysis at IBM Research. "Analog designs have used statistical methods for years."

Visweswariah said EinsTimer can handle front-end variations like L-effective, hot electron effect and different voltage thresholds; back-end variations such as metal separation and width; and environmental variations like temperature and Vdd.

What most distinguishes statistical timing analysis is the outputs it produces. They may include probability distributions for slack margins, sensitivity reports that link timing and process variables, predictions of worst-case performance across the performance space, and yield curves for different frequencies.

New mind-set
The problem is the input. Statistical timers need to know the sensitivity of the delay of library elements to each process variable. EinsTimer comes with characterization software that can extract that information from foundry-supplied libraries, Visweswariah said.

Visweswariah acknowledged that statistical timing requires a new way of thinking, since what used to be one number now becomes a probability distribution. But he said there are ways of easing into statistical analysis, such as asking for a Three Sigma report on slack distribution.

Although the statistical capability is built on top of the EinsTimer static timing analyzer, designers can continue to sign off with other static timing analyzers, such as Synopsys' market-leading PrimeTime, Visweswariah said.

Although the suite is available now, IBM declined to discuss pricing.