The concept of ultra-low-power CMOS technology involves a substantial reduction of the supply and threshold voltages to reduce the power dissipation. Going to very small voltages (well below 1V) technology evaluation and optimization cannot just rely on traditional device parameters but must be done at least on the circuit level. This requires very accurate device models for circuit simulation.
As a consequence of the low supply and threshold voltages ULP devices operate in the transition region between weak and strong inversion, i.e., moderate inversion, which becomes predominant for the device behavior. Unfortunately though, this is a major weak point in most existing compact models. Similar problems arise when future VLSI technologies are to be investigated and the existing compact models are not yet refined to cover all relevant effects accurately.
The interpolation of terminal currents seems attractive, but very often the accuracy in the weak-inversion regime is insufficient. Early work in this field of so-called table look-up MOSFET modeling started with timing simulation [14,25], extending to general circuit simulation [70,71,19] and special techniques for scalability and data compression [30,31]. While these models serve well for strong inversion, no methods were given to handle the moderate and weak-inversion regime with high accuracy. Capacitance modeling is essentially restricted to analytical models.
This work follows a new approach based on a special interpolation of terminal currents and charges with physically motivated functions, i.e., piecewise polynomial and/or exponential splines. This method ensures that various physical effects are accounted for implicitly. Using a simple transient method for device characterization the input data for the model can be obtained from measurements or simulations without any a priori knowledge and without any parameter fitting. The primary aim of the model is to render the behavior of a given device in a direct way to a circuit simulator to obtain accurate circuit performance data, which can then be related to process parameters. Consequently, physical interpretability is abandoned in favor of accuracy and generality of the model.