You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!

  • How accurate are AMS Behavioral Models?

    You often hear the mantra that Analog Mixed-Signal (AMS) behavioral models are inaccurate. It is repeated so often that it starts to morph into “AMS models are inherently inaccurate”. This is not true.
    Why do RTL designers need to care about timing constraints?-op-amp-symbol-jpg

    Source: Wikipedia

    The accuracy of a model is determined only by what the modeler has decided to include in the model.

    This, of course, is obvious – and true for any model. But it is forgotten when it is assumed that AMS behavioral models are inherently inaccurate.

    There is no doubt that most AMS behavioral models have been designed to be “inaccurate” – in an attempt to:

    1. Save simulation time
    2. Make them easy to use, understand, and, often, to create.

    If an AMS behavioral model is designed to imitate the terminal behavior of a device or block, can it be as accurate, or even more accurate, than a SPICE-level model that simulates all the components inside?

    The answer is “yes”. Especially when the external behavior results from an aggregation of many contributors, each of which needs to be modeled accurately in a SPICE-level model to get the terminal behavior correct. For example, the offset voltage drift (variation with temperature) of an op amp has many contributors, some more important than others. A SPICE-level implementation would have to correctly model all of the temperature variations of all of the offset voltage contributors. A behavioral model simply fits a curve through the measured variation of offset voltage with temperature – and is therefore inherently as accurate as the fit.

    Admittedly, this is a characteristic that is rarely important – it is just used here to prove the point. A better example that shows how a behavioral model can be extremely accurate is the op3 op amp model developed by Ken Ruan and myself at Analogy (now part of Synopsys) [1]. Op3 contains, on top of the typical op amp characteristics of gain, open-loop bandwidth, slew rate, input bias currents etc, CMRR, no limits on the poles and zeros, PSRR, common-mode rejection range and latch-up, crossover distortion, non-symmetric and non-linear output characteristics, balance adjust, noise, voltage gain and slew rate variation with supply, no internal ground node and temperature effects. As well, it uses the inherent behavioral modeling ability to change configuration by simply changing parameters: it has parameters that automatically change the configurations for BJT, MOS, JFET or transimpedance input stages, external or internal compensation, GaAs or silicon and 8 different output stage configurations.

    There is no doubt that most AMS behavioral models are designed to be less accurate than SPICE-level models. But that’s a conscious decision (hopefully) made by the modeler when determining the accuracy/simulation time tradeoff. It is not an inherent limitation of AMS behavioral models.

    [1]. Op3: A Behavioral Generic Model of Operational Amplifiers, Chapter 4 in the book Analog and Mixed-Signal Hardware Description Languages, eds. A Vachoux, J-M Berge, O. Levia & J. Rouillard, Kluwer Academic Publishers, 1997

    Why do RTL designers need to care about timing constraints?-20nm-ref-flow-banner-gif