next up previous contents
Next: 6.4 Optimization Process Up: 6. Gate Delay Time Previous: 6.2.2 Infinite Inverter Chain


6.3 Simulator Setup

In contrast to the drive current optimizations, the MINIMOS-NT input deck for the gate delay time optimizations is much more complex because it contains additional functionality to control the transient simulations and to support the input/output file handling. Nevertheless, there is still a single input deck that is used for both the on- and off-transitions, again distinguished by the keyword onoff.

Considering the highly increased computational effort for two-dimensional transient simulations of a complete inverter stage compared to the DC simulations of only one MOS transistor, all possible ways to keep the overall simulation time short have to be exploited.

The simulator setup utilizes mixed-mode transient simulations using basic drift-diffusion equations. This time the current density of the minority carriers (holes in case of the NMOS, electrons in case of the PMOS transistor) cannot be neglected as they play an important role in charging and discharging the space charge regions of the MOS transistor. The gate work function differences are set to $-$0.55 eV for the NMOS and 0.55 eV for the PMOS transistor assuming that a double gate polysilicon process was used.

The repeat level handling for evaluation steps is the same as for the drive current optimizations (see Table 4.1): First the initialization file is deactivated, then the direct Gauss-solver is tried instead of the linear solver, and then the termination criterion is loosened.

Initialization files are basically used for every simulation, but in case of an evaluation step, as shown in Fig. 6.5, only for the first time-step. In the case of a gradient step the initialization files are used for every time-step. This speeds up the simulations since the initialization files result from a previous evaluation step and contain results that are already very close to the final solution of the gradient simulations. This is usually not the case for an evaluation step where a file initialization for every time-step would slow down the simulations.

Additionally, a good initial guess is provided for the output node of the inverter which is 0 V for the on-transition and $V_{\mathrm{dd}}$ for the off-transition. A transient simulation of the inverter is usually finished when the output voltage crosses the $V_{\mathrm{dd}}$/2 mark which is automatically detected in the MINIMOS-NT input deck. Only for every last evaluation step before the gradient calculations (the temporary minima shown in Fig. 6.5) the simulation lasts longer, until the output voltage settles to 5% of the final voltage. This is necessary because the input I-t and output V-t curves of this run will be used for the load C-V and input V-t curves evaluations, as explained in Section 6.2. Therefore, they have to cover a wider range than for the delay time evaluations.

A sophisticated time-step control is employed for the transient simulations. Because the accuracy of the output curve has to be high at the point where the delay time is extracted (at $V_{\mathrm{dd}}$/2), the actual delay time is guessed from the delay time of the previous local optimum, and a time-step refinement is utilized at this particular point.

Fig. 6.6 shows how this works: The current time-step $s$ is decreased starting from $s_{\mathrm 0}$ at t = 0 s until the output voltage reaches $V_{\mathrm{dd}}$/2. At that time the time-step reaches its minimum value of $s_{\mathrm 1}$. Then it starts to increase again. The time-steps are marked with dots in Fig. 6.6.

This time-step behavior is implemented using the following equation:

\begin{displaymath}
s(t) = \left( s_{\mathrm 0}^{-1} + s_{\mathrm 1}^{-1} \cdot ...
...- t_{\mathrm d,in})^2}{2 \cdot \sigma^2} \right)
\right) ^{-1}
\end{displaymath} (6.4)

It utilizes a Gaussian function with standard deviation $\sigma$ to provide a smooth transition between the start time-step $s_{\mathrm 0}$ and the refinement time-step $s_{\mathrm 1}$. The following parameters are chosen for the refinement control and have delivered good results:

\begin{displaymath}
s_{\mathrm 0} = \frac{t_{\mathrm d,in}}{4}
\end{displaymath} (6.5)


\begin{displaymath}
s_{\mathrm 1} = \frac{t_{\mathrm d}}{10}
\end{displaymath} (6.6)


\begin{displaymath}
\sigma = 6\cdot s_{\mathrm 1}
\end{displaymath} (6.7)

where $t_{\mathrm d}$ is the guess for the delay time received from a previous optimization step (44 ps in this case) and $t_{\mathrm d,in}$ is the input delay time which has a constant value of 50 ps.

Figure 6.6: The time-step control method for the transient simulation of the inverter stage transition.
\resizebox{0.95\textwidth }{!}{
\psfrag{xlabel} [ct][ct] {time $t$\ (ps)}
\psfra...
...udegraphics[width=0.95\textwidth ]{../figures/timesteprefinement-gatedelay.eps}}


next up previous contents
Next: 6.4 Optimization Process Up: 6. Gate Delay Time Previous: 6.2.2 Infinite Inverter Chain
Michael Stockinger
2000-01-05