next up previous contents
Next: 4. Device Modeling for Up: 3.3 VLSI Performance Metric Previous: 3.3.7 Noise Margins, Inverter


3.3.8 Application

To obtain the raw device data MINIMOS [27] is used as device simulator and the device structures can be generated either by coupled process simulation, or with an additional program from analytical profile descriptions.

Table 3.6 shows results from a pocket-implanted LDD NMOS device with $\ensuremath{L_{\mathit{nom}}}\xspace = 0.18$$\mu$m, $\ensuremath{t_{\mathit{ox}}}\xspace = 5$nm, and $\ensuremath{X_{\mathit{j}}}\xspace = 0.09$$\mu$m designed for operation at $\ensuremath{V_{\mathit{DD}}}\xspace = 1.5$V. The system parameters used for analysis were $\ensuremath{{\mathit{ld}}}\xspace = 7$ and $\ensuremath{{\mathit{ar}}}\xspace = 0.1$ (interconnect capacitances were not considered). The evaluation was carried out for the nominal case ``nominal'' and two worst-case corners (``a'': $\ensuremath{T}\xspace =\ensuremath{T_{\mathit{max}}}\xspace =125$$^{\circ }$C, channel and pocket implant doses, and $\ensuremath{L}\xspace $ reduced by 20%; and ``b'': $\ensuremath{T}\xspace =\ensuremath{T_{\mathit{min}}}\xspace =0$$^{\circ }$C, channel and pocket implant doses increased by 50%, and $\ensuremath{L}\xspace $ increased by 20%). The data ``c'' are the same as ``a'' except that $\ensuremath{t_{\mathit{ox}}}\xspace $ was reduced instead of the channel and pocket implants (to obtain the same shift in $\ensuremath{V_{\mathit{T,lin}}}\xspace $). Each of these evaluations takes a CPU time of about 5 minutes on an HP735 workstation. The errors of the inverter gain and noise margins determined from device-level simulations of an inverter are typically in the order of 5%.


Table 3.6: Performance of a pocket-implanted LDD NMOST ( $\ensuremath {L}\xspace =\rm0.18\mu m$, $\ensuremath{V_{\mathit{DD}}}\xspace =\rm1.5V$)
case $\ensuremath{V_{\mathit{T,lin}}}\xspace $ $\ensuremath{t_{\mathit{d}}}\xspace $ 1 $\ensuremath{E_{\mathit{s}}}\xspace $ $\ensuremath{f_{\mathit{c,max}}}\xspace $ $\frac{\ensuremath{f_{\mathit{c,max}}}\xspace }{\ensuremath{f_{\mathit{c,min}}}\xspace }$ $\ensuremath{{\mathit{NM}}}\xspace $ $\ensuremath{A_{\mathit{inv}}}\xspace $
  [V] [ps] [fJ] [GHz]   [Vdd]  
nominal 0.44 47.6 1A.1 3.0 A.5e4 0.36 B.7
a: \bgroup\color{red}$\textcolor{red} {N_{ch}\downarrow T\uparrow L\downarrow}$\egroup 0.21 34.2 27.2 4.2 7.3e0 0.29 6.1
b: \bgroup\color{blue}$\textcolor{blue} {N_{ch}\uparrow T\downarrow L\uparrow}$\egroup 0.67 B2.6 1B.3 1.5 2.6e9 0.44 34.2
c: \bgroup\color{orange}$\textcolor{orange}{\ensuremath{t_{\mathit{ox}}}\xspace \downarrow T\uparrow L\downarrow}$\egroup 0.23 35.9 1B.8 4.0 7.1e1 0.32 B.9
error           -2.7% 3.6%
1 k1=0.67, k2=2.5, k3=1 (CMOS with slow PMOST)

The curves in Figs. 3.8-3.11 are the delay time, switching energy, maximum clock frequency, and clock frequency margin   $\ensuremath{{f_{\mathit{c,max}}/f_{\mathit{c,min}}}}\xspace $ for the three cases. From Fig. 3.11 and Table 3.6 it can be seen that the device fails at corner ``a'' because $\ensuremath{{f_{\mathit{c,max}}/f_{\mathit{c,min}}}}\xspace $ is too small for dynamic logic. The noise margins, however, are still 30% $\ensuremath{V_{\mathit{DD}}}\xspace $, which would allow static-logic operation.

Figure 3.8: Inverter delay \ensuremath{t_{\mathit{d}}} vs. \ensuremath{V_{\mathit{DD}}}
\includegraphics[scale=1.0]{lp-td.eps}

Figure 3.9: Switching energy \ensuremath{E_{\mathit{s}}} vs. \ensuremath{V_{\mathit{DD}}}
\includegraphics[scale=1.0]{lp-es.eps}

Figure 3.10: Maximum clock frequency \ensuremath{f_{\mathit{c,max}}} vs. \ensuremath{V_{\mathit{DD}}}
\includegraphics[scale=1.0]{lp-fc.eps}

Figure 3.11: Clock frequency margin $\frac{\ensuremath{f_{\mathit{c,max}}}\xspace }{\ensuremath{f_{\mathit{c,min}}}\xspace }$ vs. \ensuremath{V_{\mathit{DD}}}
\includegraphics[scale=1.0]{lp-fm.eps}

Figure 3.12: Drive current \ensuremath{I_{\mathit{on}}} vs. \ensuremath{V_{\mathit{DD}}}
\includegraphics[scale=1.0]{xx-ion.eps}

Figure 3.13: Leakage current \ensuremath{I_{\mathit{off}}} vs. \ensuremath{V_{\mathit{DD}}}
\includegraphics[scale=1.0]{lp-ioff.eps}

The impact of supply and threshold voltage on maximum clock frequency and switching energy is shown in Figures 3.14 and 3.15. For the simulations $\ensuremath{t_{\mathit{ox}}}\xspace $ was varied from 1nm to 18nm, and $\ensuremath{f_{\mathit{c,max}}}\xspace $ and $\ensuremath{E_{\mathit{s}}}\xspace $ were plotted over $\ensuremath{V_{\mathit{DD}}}\xspace $ and the extracted $\ensuremath{V_{\mathit{T,lin}}}\xspace $. To find an optimum device, one can, e.g., pick a contour in Fig. 3.14 according to the required performance and then, along this contour, find the minimum $\ensuremath{E_{\mathit{s}}}\xspace $ in Fig. 3.15. In comparison, taking the product $\ensuremath{E_{\mathit{s}}}\xspace \ensuremath{t_{\mathit{d}}}\xspace $ as an optimization criterion in Fig. 3.17 would favor devices with a close-to-optimal power efficiency. Most notably, the optimum \ensuremath{V_{\mathit{DD}}} lies in the range of 0.2...0.4V although these devices were designed for $\ensuremath{V_{\mathit{DD}}}\xspace =\rm1.5V$. Figures 3.18 and 3.19 show the same plots of \ensuremath{f_{\mathit{c,max}}} and \ensuremath{E_{\mathit{s}}} for dedicated Ultra-Low-Power devices designed for $\ensuremath{V_{\mathit{DD}}}\xspace =\rm0.2V$ (process A in Section 2.5)

Figure 3.14: Maximum clock frequency \ensuremath{f_{\mathit{c,max}}} vs. \ensuremath{V_{\mathit{DD}}} and \ensuremath{V_{\mathit{T,lin}}}
\includegraphics[scale=0.9]{lpx-fc.eps}

Figure 3.15: Switching energy \ensuremath{E_{\mathit{s}}} vs. \ensuremath{V_{\mathit{DD}}} and \ensuremath{V_{\mathit{T,lin}}}
\includegraphics[scale=0.9]{lpx-es.eps}

Figure 3.16: $\ensuremath{f_{\mathit{c,max}}}\xspace /\ensuremath{f_{\mathit{c,min}}}\xspace $ vs. \ensuremath{V_{\mathit{DD}}} and \ensuremath{V_{\mathit{T,lin}}}
\includegraphics[scale=0.9]{lpx-fm.eps}

Figure 3.17: 1/( \ensuremath{E_{\mathit{s}}} \ensuremath{t_{\mathit{d}}}) vs. \ensuremath{V_{\mathit{DD}}} and \ensuremath{V_{\mathit{T,lin}}}
\includegraphics[scale=0.9]{lpx-pf.eps}

Figure 3.18: Maximum clock frequency \ensuremath{f_{\mathit{c,max}}} vs. \ensuremath{V_{\mathit{DD}}} and \ensuremath{V_{\mathit{T,lin}}} (0.2V ULP technology)
\includegraphics[scale=0.9]{ulp-fc.eps}

Figure 3.19: Switching energy \ensuremath{E_{\mathit{s}}} vs. \ensuremath{V_{\mathit{DD}}} and \ensuremath{V_{\mathit{T,lin}}} (0.2V ULP technology)
\includegraphics[scale=0.9]{ulp-es.eps}

This new approach combines the immediate relevance of system models with the accuracy of device simulation at minimum computing effort. Yielding system information directly for arbitrary operating points, like worst-case process corners, makes this method an ideal tool for fast device evaluation without the need of full compact-model characterization during the optimization process.


next up previous contents
Next: 4. Device Modeling for Up: 3.3 VLSI Performance Metric Previous: 3.3.7 Noise Margins, Inverter

G. Schrom