Nano Design Needs New Modeling

Electronic News, April 2, 2001 by John Croix

AUSTIN--At 150 nanometers and below, subtle physical effects invalidate conventional macro and cell models that form the basis for delay calculations that are fundamental to physical design methods. As a result, chip performance and delivery schedules slip as designers attempt to increase guardbands and achieve timing sign-off. Using more dynamic approaches that combine data with dynamic analysis algorithms, new modeling methods based on instance-specific operating points (ISOPs) are emerging as the preferred modeling approach for nanometer physical design.

Many years ago designers worked with cell-delay models that were measured as constant multipliers against some standard-delay value. As feature sizes shrank, delay models became more sophisticated--moving from a simple slope/intercept equation of a line to more advanced forms including table lookup and multi-coefficient polynomials. Interconnect delay and slew followed a similar evolution. Interconnect values evolved from constants, through wire load tables, and into pole/residue forms in order to achieve near-SPICE accuracy.

While timing is still a large concern for most designs, other very deep-submicron (VDSM) effects now play equally important roles. Signal-integrity, noise, process variation, over-the-cell routing and metal migration, just to name a few, have increased the complexity of the analysis. These types of analysis tools have driven the advancement of cell and interconnect models. Models are no longer the monolithic data elements of the past. Instead, values are measured across a multidimensional space in very fine steps of resolution. With more data to draw upon, analysis algorithms are given greater flexibility and can produce numbers that are more accurate than previously achievable.

Active Models

We now stand at the cusp of the next evolutionary point in modeling. Today, we see a migration from data-only models to models that combine data and algorithms together for greater accuracy. Just as programmers moved from the separate data/algorithm spaces provided by traditional languages (such as C and Pascal) to object-oriented languages (such as C++) that combine data and algorithms, CAD models need to evolve to a data-plus-algorithm form.

Fortunately a data-plus-algorithm model already exists in the form of Open Library API (OLA), invented by IBM Corp., promoted by SI2 (www.si2.org), and standardized by the Institute of Electrical and Electronics Engineers (IEEE 1481). Commercial OLA offerings are now emerging to provide, we believe, a fast track to the OLA modeling paradigm.

Data-only models are those models that are generated via numerical processing, placed into a static form, and whose eventual result relies upon the proper setting of some variable(s). An application is responsible for reading the data that represents the model, setting known values (required for the evaluation of the model), and applying an algorithm within the application to evaluate the model. In other words, the algorithm to evaluate the model is contained within the application, not within the model.

Active models provide a new degree of control to the model writer. An active model can be as simple as a data-only model plus the algorithm to evaluate the data-only model. An active model can also be quite sophisticated and may include if-then-else branches, for-loop constructs, environmental detection and instance-specific isolation to determine the response of the model.

Another benefit of active models is consistency. In a data only model, each application provides the algorithm to evaluate the model. Thus, if two applications treat a data-only model differently (for example, one might zero negative numbers while the other uses the negative values), timing closure may be impossible to achieve without additional guardbanding or additional iterations, forestalling timing sign-off and time-to-market. Active models also represent the new way of thinking about the manner in which cells/macros and interconnects behave and interact. Active models can interact with one another to more accurately determine the cell/macro and interconnect delay and slew rates. An active model can determine its environment, calculate the effect of an active load and return the correct result.

ISOP Behavior

ISOP refers to the behavior of a specific instance of a model within a design. Consider an active model that uses temperature in the calculation of cell intrinsic delay and output slew rates. In a static-timing program you might set the global temperature of a design and generate static-timing reports. However, these reports are only valid for the temperature specified across the entire design and do not take into account regional variations.

Active models are the next great step forward in the cell and interconnect modeling paradigm. By merging the data with the algorithm to evaluate the model, designers will see a greater degree of consistency between disparate applications used in the flow. Guardbanding can be reduced since the models will no longer be force-fit into outdated fixed form equations or tables. And, through the use of techniques such as instance-specific operating point analysis, models can be evaluated in an environment that better represents their true operating conditions rather than the conditions imposed by any single application.