Alpert/Handbook of Algorithms for Physical Design Automation AU7242_C041 Finals Page 872 24-9-2008 #9 872 Handbook of Algorithms for Physical Design Automation 41.4.1 EFFECTS OF INDUCTANCE ON DELAY AND SIGNAL RISE TIME A general expression for the propagation delay from the input to the output of an RLC line of length l with an ideal power supply and an open circuit load is given by [18] t pd = √ LC e −2.9 ( α asym l ) 1.35 l + 0.74α asym l 2 (41.8) where α asym = R 2 C L (41.9) α asym is the asymptotic value at high frequencies of the attenuation per unit length of the signals as the signals propagate across a lossy transmission line, as shown in Figure 41.4. For the limiting case where L → 0, Equation 41.8 reduces to 0.37RCl 2 , illustrating the square dependence on the length of an RC wire as aforementioned.For the other limiting case where R → 0, the p ropagation delay is given by √ L t C t = l √ LC. Note the linear dependence on the length of the line. Note also that inductance always increases the delay as compared to an RC mod e l; i.e., if inductance is neglected, the delay is underestimated by the incomplete RC model. The rise time of signals propagating across RLC lines improves as the inductance effects of the line increase. This behaviorcanbe explainedby referring toFigure41.4, which depictsthe attenuation of signals as they travel across an RLC line as a function of frequency.Higher frequency components at the edges of a pulse suffer greater attenuation as compared to low frequency components. The shape of a signal degrades as the signal travels across a lossy transmission line because of the loss of these high-frequency components. The attenuation constant becomes less frequency dependent as inductance effects increase or as R/ωL decreases as shown in Figure 41.4. In the limiting case of a lossless line representing maximum inductance effects, the attenuation constant α is 0. Thus, as inductance effects increase, a pulse propagating across an RLC line maintains the high-frequency components in the edges, improving the signal rise and fall times. A t t e n u a t i o n c o n s t a n t a R = 4 0 0 Ω / c m R = 1 0 0 Ω / c m R = 5 0 Ω / c m R = 1 0 Ω / c m R = 2 0 0 Ω / c m 2 1 0 0 2 ϫ 1 0 9 4 ϫ 1 0 9 6 ϫ 1 0 9 8 ϫ 1 0 9 1 ϫ 1 0 1 0 F r e q u e n c y ( H z ) FIGURE 41.4 Attenuation constant versus frequency. L = 10 nH/cm, C = 1pF/cm, and R is 10, 50, 100, 200, and 400 /cm, respectively. Alpert/Handbook of Algorithms for Physical Design Automation AU7242_C041 Finals Page 873 24-9-2008 #10 Inductance Effects in Global Nets 873 41.4.2 EFFECTS OF INDUCTANCE ON POWER DISSIPATION Power consumption is an increasingly important design parameter with mobile systems and high performance, high-complexity circuits such as leading edge microprocessors. If the frequency of switching is f cycles per second, then the dynamicpower consumption is describedby thewell-known formula, P dyn = C t V 2 DD f (41.10) The dynamic power depends only on the total load capacitance, the supply voltage, and the operating frequency. As discussed in Section 41.7, increasing inductance effects result in a lower number of repeaters as well as smaller repeater size. The smaller size and number of repeaters therefore signif- icantly reduces the total capacitance of the repeaters and, consequently, reduces the total dynamic power consumption. The short-circuit power results from the NMOS and PMOS b locks o f a CMOS gate being on simultaneously during the rise and fall times of the input signal, creating a current path between the power supply and ground. As discussed in Section 41.4.1,the inductance reduces the rise time of the signalsin an integrated circuit, reducing theshort-circuitpower.The short-circuit power consumption of a gate driven by an RLC line versus the line inductance is plotted in Figure 41.5. Note that as inductance effects increase, the short-circuit power consumption significantly decreases due to the faster input rise time. Also, the smaller repeater sizes dramatically reduces the short-circuit power consumption because the short-circuit power of a CMOS gate is super linearly dependent on the transistor widths. Finally, it has been shown in Ref. [19] that the short-circuit power consumption of a CMOS gate decreases as the inductance of the driven net becomes more significant. Intuitively, inductance is an element that does not consume any power while resistance consumes power. Hence, as the interconnect behavior becomes dominated by inductance rather than resistance, the power consumption of integrated circuits will be redu ced. 41.4.3 EFFECTS OF INDUCTIVE COUPLING ON DELAY UNCERTAINTY In a set of inductively and capacitively coupled lines, the signal propagation delay on a particular line reaches a minimum when neighbor lines are switching in the same direction. The delay on that line reaches a maximum when that particular line is switching in opposite direction to neighbor lines because of the increased effective capacitance that has to be charged or discharged. The ratio of Short-circuit power (pJ/cycle) 2.5 2 1.5 1 0.5 0 024 6 81012 L t (nH ) FIGURE 41.5 Short-circuit energy consumed per cycle by CMOS gate driven by an RLC line versus the inductance of the line. The total resistance and capacitance of the line are maintained constant at 100 and 1 pF, respectiv ely. Alpert/Handbook of Algorithms for Physical Design Automation AU7242_C041 Finals Page 874 24-9-2008 #11 874 Handbook of Algorithms for Physical Design Automation the maximum and minimum signal delays on a certain signal line can be defined as d elay uncertainty (D U ) for that line, as given by D U = t d max t d min (41.11) For RLC lines, the delay is given by [20] t d = 1.047 ·E · τ LC + 1 ·4 ·τ RC (41.12) where E is a term that depends on the damping factor of this line as described in Ref. [20], τ RC = k R k · r,j C rj ·(α r − α j ) (41.13) and τ LC = k L k · r,j C rj · (α r − α j ) +M k · l,m C lm · ( α l − α m ) (41.14) Each line has a switching factor associated with it and is d enoted α i for interconnect i. The switch- ing factor takes the values 1, 0, and −1 for lines switching from low-to-high, nonswitching lines, and lines switching from high-to-low, respectively. k runs over all the branches on the path from the primary input to node i onthetree(whichi belongs to), r runs over all the nodes downstream of k on that tree, and j runs over all the nodes to which r has a capacitance connected to. In the case of capacitances to ground, j = 0. The index l runs over all the nodes downstream of M k on the coupled tree (which i does not belong to). The index m runs over all the nodes, which l has a capacitance connected to. The time constants τ RC and τ LC depend on the switching directions of neighborlines. As neighbor lines switch in opposite directions to the line in consideration, τ RC is maximum. When neighborlines switch in the same direction, τ RC is minimum as given by Equation 41.13. On the other hand, τ LC decreases when neighbor lines switch in opposite directions because of the opposite currents in neighbor lines, which causes negative mutual inductance terms to appear in Equation 41.14. When neighbor lines switch in the same direction, τ LC increases because the mutual inductances add to the self-inductance as in Equation 41.14. This opposite behavior of τ RC and τ LC results in reducing the discrepancy between the maximum and the minimum delays of a line because of coupling with other lines. Circuit simulation(Figure41.6)for the signal on the middle line of three coupledlines showsthat as inductance effects increase, the ratio between the maximum and the minimum delays decreases. That is, higher inductive effects lead to lower d elay uncertainty and narrower switching windows. Lowering delay uncertainty is a positive effect of inductance because narrower switching windows give significant degrees of freedom in physical design to limit noise and control glitches among many other benefits. 41.5 INDUCTIVE NOISE As discussed before, aline can inductively couple to lines thatare far away unlikecapacitivecoupling, which only occurs between adjacent lines. The problem of inductive coupling is particularly severe in wide busses, which are commonplace in most digital integrated circuits such as DSP and micro- processor circuits. The width of busses in digital circuits is continuously increasing with technology Alpert/Handbook of Algorithms for Physical Design Automation AU7242_C041 Finals Page 875 24-9-2008 #12 Inductance Effects in Global Nets 875 Delay uncertainty Minimum delay Maximum delay 0 50 100 150 200 250 Delay (ps) Increasing inductance FIGURE 41.6 Delay uncertainty dependence on inductance effects. The data shown are for the delay of the middle line of three coupled parallel lines. scaling. Hence, the problem of inductive coupling in busses will have even more significance in future technologies. Physically, a wide bus with all the lines switching in the same direction behaves as one wide line. Such a wide line has much higher inductance effects as compared to any of the individual lines in the bus. Hence, the effective inductance of a line that is part of a bus is far larger than the self- inductance of that line. This fact can also be quantitativelyunderstoodbyreferringto Equation 41.14, which shows that if all the lines are switching in the same direction, the LC time constant of the line becomes much larger than the case of an individual line because of all the mutual terms adding to the self-inductance term. This increase in the LC time constant means much higher overshoots and inductive noise on any line in the bus. To examine theimpact of inductance oncircuit cross talk in ahigh-performance 0.18-µm process, the worst-case noise generated on an 8-bit, 3000-µm long, standard data bus was simulated in Refs. [9,21]. The bus was implemented in metal 6 with all lines having a metal width of 3 µmanda metal-to-metal spacing of 1.5µm consistent with typical high-level metal implementations of high- performance globalbusses. Thedata buswasalso sandwichedb etween aV DD line anda GNDline each 15-µm wide to provide a return path for the current flowing in the buses. The drivers and receivers were implemented using simple buffers. A distributed RLC model for the interconnect was produced where FastCap [22] was used to model the interconnect capacitance, and FastHenry [15] was used to model both the resistance and the inductance of the interconnects. By applying a 5-ps rise time step signal to all the inputs except the one in the middle, SPICE simulations show a totally unacceptable voltage glitch of 1.17 V. Such a glitch could cause erroneous switching and logic failures. Note that if the inductance is neglected and not modeled, the cross-talk noise becomes only 0.59 V, which is almost half the value of the actual glitch. This shows the importance of modeling on-chip inductance for accurate detection of cross-talk voltage glitches. In terms of substrate coupling, inductance effects increase this type of noise significantly. Over- shoots and undershoots owing to inductance cause noise coupling through the common substrate, which is both difficult to measure and difficult to control. Substrate noise-conduction modes can be classified into (1) resistive coupling, (2) capacitive coupling, (3) impact ionization, and (4) body effect (Figure 41.7). All these modes involve currents running into the substrate from the drains or the sources of transistors and affecting other devices. Ideally, the p-bulk is grounded, which always Alpert/Handbook of Algorithms for Physical Design Automation AU7242_C041 Finals Page 876 24-9-2008 #13 876 Handbook of Algorithms for Physical Design Automation Contact Contact n-well n + n-well p+ source p+ drain n + drain n + source n + bulk p+ bulk Impact ionization p-epitaxy Capacitive coupling Negative transient causing depletion layer increase Resistive coupling FIGURE 41.7 Mechanisms of substrate noise propagation in an integrated circuit. reversebiases the p−n junctionsat the drains and sources of NMOS transistors assuming that ground is the lowest voltage th at can appear at the drain o r source of any transistor. Similarly, the n-well is connected to V DD to reverse bias the drains and sources of PMOS transistors. However, inductance effects cause overshoots and undershoots that can forward bias these junctions resulting in currents flowing into the substrate causing substrate coupling. For that reason, substrate coupling noise is sometimes called bootstrap noise. Also, if the bulk is biased with a switching ground bus, the ground on the bulk is not perfectbecause switching transients will cause voltage dropsacross the line. Hence, the switching transients on the power supply line can couple to transistors resistively through the p+ bulk contacts. The parallel summation of bulk contacts and epitaxy resistances provides a very low impedance path (nearly short) to the p+ buried layer . The second source of substrate noise is capacitive coupling through the MOSFET source and drain p − n junctions. Each n-well on p+ bulk also introduces fairly large p − n junctions forming a capacitance b etween the V DD rails biasing the n-well and the V SS rails biasing the bulk. The noise injected into the substrate via capacitive coupling is inversely proportional to the rise time of the signals on the drains and sources of transistors. As discussed in the previous section, inductance effects result in faster signal transition times. Another source of substrate noise is impact ionization current, generated at the p inch-off point of the NMOS transistors. Impact ionization causes a whole current inthe bulk.A negativebulk transientwill increasethe depletion region between thesource and the bulk. This depletes the channel of charge carriers and increase V th . The total effect is a sporadic decrease in the I DS current. In general, these transients increase with higher inductance effects owing to the higher voltage swings and overshoots. 41.6 REQUIREMENTS ON CAD TOOLS AND THEIR PERFORMANCE The signals th at occur in RLC circuits are significantly more complicated than signals in RC circuits. For example, the RLC signals shown in Figure 41.1 have overshoots, very large inertial delay, fast rise time, and are rich in harmonics. Hence, new delay models and model order reduction techniques are required to handle RLC circuits. One of the most popular approximate delay models used for the design and analysis of integrated circuits is the Elmore delay model. This first order delay model cannot be used with RLC circuits with underdamped responses, because underdamped responses involve complex poles that appear in conjugate pairs. Hence, at least a second order approximation is required for RLC circuits. One such model was developed in Ref. [17] and maintains the popular characteristics of Elmore delay. Model order reduction techniques allow the calcu lation of approximations of higher orders to accurately simulate the interconnect. Asymptotic waveform evaluation (AWE) is one popular tech- nique used successfully with RC interconnects [23]. However, AWE cannot calculate enough poles Alpert/Handbook of Algorithms for Physical Design Automation AU7242_C041 Finals Page 877 24-9-2008 #14 Inductance Effects in Global Nets 877 to handle complex underdamped responses because of numerical errors. Hence, a set of new model order reduction techniques have b een recently developed that are capable of calculating h igher order approximations necessary for simulating systems with complex responses. Examples are Pade via Lanczos (PVL) and matrix Pade via Lanczos (MPVL) [24], Arnoldi algorithms [25], block Arnoldi algorithms [26], p assive reduced-order interconnect macromodeling algorithm (PRIMA) [27], and SyPVL algorithm [28]. However, these model order reduction techniques can have significantly higher computational complexity than AWE. Hence, there is a need for innovative simulation techniques for handling complex responses arising in RLC circuits. Another feature that complicates the analysis and design of integrated circuits including on-chip inductance, is the far-reaching inductive coupling to other lines in the integrated circuit. Typically, a line in a wide bus couples to all the lines in the bus. As compared to capacitive coupling, which only couples a line to the immediate neighbors, inductive coupling results in larger circuits (the whole bus rather than three lines) to be analyzed, and these circuits have a significant amount of inputs. In general, all CAD tools will run significantly slower when using an RLC model as compared to an RC model. This behavior is simply due to the more complex model used and the higher signal integrity issues involved. In addition to the lower performance of CAD tools, a very large infrastructure o f RC-based CAD tools needs to be modified to include inductance effects. 41.7 PHYSICAL DESIGN INCLUDING INDUCTANCE EFFECTS Currently, the industry applies a three-step design process for integrated circuits when handling inductance. First, employ design methodologies and techniques to reduce the inductance effects in the design. Second, use the well-developed RC-based design tools to optimize and verify the circuit. Third, pray nothing will go wrong. However, as discussed in this chapter, inductance can have useful effects such as improving the rise time of signals, reducing the power consumption, and reducing the number o f inserted repeaters. Hence, by suppressing inductance effects and using RC-based tools, a suboptimal circuit results in terms of area, power consumption, and speed. Also, signal integrity and cross-talk issues owing to inductive coupling are neglected, which can result in undetected reliability problems. Fortunately, in recent years many researchers h ave started mod ify ing design methodologies and physical design to include inductance rather than suppressing it. A sample of these works is discussed in this section. Including inductance in interconnect routing has been dealt with in several works. An example is the work in Ref. [29]. The work mainly deals with reducing capacitive and inductive cross talk within the interconnect during full chip routing. The work shows a reduction in cross talk by a factor of 2.5 while increasing the routing area with less than 5 percent as compared to an algorithm that does not include cross talk. In wide busses, inductive cross talk can sometimes exceed capacitive coupling when all lines switch in the same d irection. Repeater insertion is another common design methodology for driving long-resistive interconnect (e.g., Refs. [3–5]). Because the RC time constant of a line is given by R t C t = RCl 2 and has a square dependence on the length of the line, subdividing the line into shorter sections by inserting repeaters is an effective strategy for reducing the total propagation delay. Currently, typical high- performance circuits have a significant number of repeaters inserted along global interconnect lines. These repeaters are large gates and consume a significant portion of the total circuit power. As discussed in Section 41.2, the amount of inductance effects present in an RLC line depends on the ratio between the RC and the LC time constants of the line. Hence, as inductance effects increase, the LC time constant dominates the RC time constant and the delay of the line changes from a quadratic to a linear dependence on the line length [18]. As a consequence, the optimum number o f repeaters for minimum propagation delay decreases as inductance effects increase. In the limit, an LC line requires zero repeater area to minimize the overall propagation delay. Inserting repeaters based on an RC model and neglecting inductance result in a larger repeater area than necessary to achieve a minimum delay. The magnitude of the excess repeater area when Alpert/Handbook of Algorithms for Physical Design Automation AU7242_C041 Finals Page 878 24-9-2008 #15 878 Handbook of Algorithms for Physical Design Automation TABLE 41.4 Total Repeater Area, Total Power, and Total Maximum Path Delay of All of the Trees Savings in Buffered RLC Savings Compared Buffered RC Totals Unbuffered Delay (percent) Model to RC (percent) Model Area (minimum inverters) 0 — 14,116 40.8 23,854 Maximum delay (ps) 6,554 42.2 3,787 6.7 4,061 Power (pJ/cycle) — — 1,379 15.6 1,632 The p ercent savings shown here represent the average savings in area, power, and maximum path delay when using an RLC model f or repeater insertion using an RC model depends upon the relative magnitude of the inductance within the RLC tree. The reduced number of inserted repeaters also simplifies the layout and routing constraints. Also, the reduced repeater area greatly reduces the power consumed by the repeaters in an integrated circuit. A more thorough analytical analysis of the effect of inductance on the repeater insertion process can be found in Ref. [18]. Practical data are listed in Table 41.4 for repeaters inserted in a large number of typical copper interconnects from a 0.25-µm CMOS technology [30]. Note that by using an RLC model rather than an RC model, a better delay can be achieved with significantly less repeater area and power consumption b y the repeaters. Inductance plays a central role in power and clock distribution networks [31–36]. Typically, the inductive characteristics of the clock distribution n etwork are intimately related to the power distribution network. The return currents from the clock determine the size of the inductive loop. These currents typ ically return in the wide, low-resistance, p ower distribution network wire s. The clock wires are typically wide enough to exhibit significant inductance effects. Full transmission line models are needed for these wires when sizing the cloc k distribution network. Several works (such as Refs. [34,35]) have dealt with these clock distribution design including inductance. In terms of the power distribution networks, inductance does dominate the total impedance of the network [31–36]. Electromigration poses an upper limit on the current density carried by the power distribution networks. Because of the unidirectional nature of the currents carried by the power distribution network, it is crucial to upsize these wires to meet the upper bound on current density. Hence, power distribution networks typically have very wide wires that exhibit significant inductance effects. However, typically a power distribution network is designed with interleaving V DD and GND lines on each layer. This design results in inductive coupling typically limited to adja- cent wires with opposing currents canceling at far distances. Hence, inductance in the power grid can be easily modeled using loop inductance models rather than partial inductances. Interestingly, designers have found ways to justify ignoring inductance in power distribution networks despite its dominance. The justification is typically that currents taken from the power distribution network are DC. There is no experimental data that supports this claim. In fact, in synchronous circuits, most of the currents are drawn around the clock edges, giving rise to very large current slopes. Several works have considered inductance in the power distribution network. However, accurate estimation of the currents drawn from the power distribution network remains crucial. REFERENCES 1. H. G. Lin and L. W. Linholm, An optimized output stage for MOS integrated circuits, IEEE Journal of Solid-State Circuits, SC-10(2): 106–109, April 1975. 2. B. S. Cherkauer and E. G. Friedman, Design of tapered buffers with local interconnect capacitance, IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 3(1): 99–111, March 1995. Alpert/Handbook of Algorithms for Physical Design Automation AU7242_C041 Finals Page 879 24-9-2008 #16 Inductance Effects in Global Nets 879 3. J. Cong, L. He, C. -K. Koh, and P. Madden, Performance optimization of VLSI interconnect, Integration: The VLSI Journal, 21: 1–94, November 1996. 4. V. Adler and E. G. Friedman, Delay and power expressions for a CMOS inverter dri ving a resistive- capacitive load, Analog Integrated Circuits and Signal Processing, 14(1/2): 29–39, September 1997. 5. S. S. Sapatnekar, RC interconnect optimization under the Elmore delay model, Proceedings of the IEEE/ACM Design Automation Conference, pp. 387–391, San Diego, CA, June 1994. 6. AS/X User’s Guide, IBM Corp., 1994. 7. Y. I. Ismail, E. G. Friedman, and J. L. Neves, Figures of merit to characterize the importance of on-chip inductance, IEEE Transactions on Very Larg e Scale Integration (VLSI) Systems, 7(4): 442–449, December 1999. 8. Y. I. Ismail, E. G. Friedman, and J. L. Neves, Inductance effects in RLC trees, Proceedings of the IEEE Great Lakes Symposium on VLSI, pp. 56–59, Ypsilanti, MI, March 1999. 9. Y. Massoud and Y. I. Ismail, On-chip inductance in high-speed integrated circuits, IEEE Cir cuits and Devices Magazine, 17(4): 14–21, July 2001. 10. Y. Massoud and J. White, Simulation and modeling of the effect of substrate conductivity on coupling inductance, Pr oceedings of the IEEE International Electron Devices Meeting, pp. 491–494, Hong Kong, China, December 1995. 11. K. Shepard and Z. Tian, Return-limited inductances: A practical approach to on-chip inductance extraction, IEEE Transactions on Computer-Aided Design, 19: 425–436, April 2000. 12. B. Krauter and S. Mehrotra, Layout based frequency dependent inductance and r esistance extraction for on- chip interconnect timing analysis, Proceedings of the IEEE Design Automation Confer ence, pp. 303–308, San Francisco, CA, June 1998. 13. A. E. Ruehli, Inductance calculations in a complex integrated circuit environment, IBM Journal of Research and Development, 16: 470–481, September 1972. 14. P. A. Brennan, N. Raver, and A. Ruehli, Three dimensional inductance computations with partial element equivalent circuits, IBM Journal of Research and Development, 23: 661–668, November 1979. 15. M. Kamon, M. Tsuk, and J. White, FastHenry: A mutipole-accelerated 3-D inductance extraction program, IEEE Transactions on Microwave Theory Technology, 42(9): 1750–1758, September 1994. 16. Y. I. Ismail and E. G. Friedman, Sensitivity of interconnect delay to on-chip inductance, Proceedings of the IEEE International Symposium on Circuits and Systems, pp. 403–407, Geneva, Switzerland, May 2000. 17. Y. I. Ismail, E. G. Friedman, and J. L. Ne ves, Equiv alent Elmore delay for RLC trees, IEEE Transactions on Computer-Aided Design, 19(1): 83–97, January 2000. 18. Y. I.Ismail andE. G. Friedman, Effects of inductance on t hepropagation delay and repeater insertion inVLSI circuits, IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 8(2): 195–206, April 2000. 19. Y. I. Ismail, E. G. Friedman, and J. L. Neves, Dynamic and short-circuit power of CM OS gates dri- ving lossless transmission lines, IEEE Transactions on Circuits and Systems 1: Fundamental Theory and Applications, CAS-46(8): 950–961, August 1999. 20. M. Chowdhury, Y. I. Ismail, C. V. Kashyap, and B. L. Krauter , Performance analysis of deep sub micron VLSI circuits in the presence of self and mutual inductance, Proceedings of the IEEE International Symposium on Circuits and Syst ems, pp. 197–200, Scottsdale, AZ, May 2002. 21. Y. Massoud, J. Kaw a, D. MacMillen, and J. White, Modeling and analysis of differential signaling for minimizing inductive cross-talk, Proceedings of the IEEE Design Automation Conference, pp. 804–809, Anaheim, CA, June 2001. 22. K. Nabors and J. White, Fast capacitance extraction of general three-dimensional structures, IEEE Transanctions on Microwave Theory Technology, 40(7): 1496–1506, June 1992. 23. L. T. Pillage and R. A. Rohrer, Asymptotic waveform evaluation for timing analysis, IEEE Tr ansactions on Computer-Aided Design, CAD-9(4): 352–366, April 1990. 24. P. Feldmann and R. W. Freund, Reduced-order modeling of large linear subcircuits via block Lanczos algorithm, Proceedings of the IEEE/ACM Design Automation Conference, pp. 474–479, San Diego, CA, June 1995. 25. M. Silveira, M. Kamon, and J. White, Efficient reduced-order modeling of frequency-dependent coupling inductances associated with 3-D interconnect structures, Proceedings of the IEEE/ACM Design Automation Conference, pp. 376–380, San Diego, CA, June 1995. 26. D. L. Boley, Krylov space methods on state-space control models, Journal of Circuits, Systems, and Signal Processing, 13(6): 733–758, May 1994. Alpert/Handbook of Algorithms for Physical Design Automation AU7242_C041 Finals Page 880 24-9-2008 #17 880 Handbook of Algorithms for Physical Design Automation 27. A. Odabasioglu, M. Celik, and L. T. Pillage, PRIMA: Passive reduced-order interconnect macromodeling algorithm, IEEE Transactions on Computer-Aided Design, CAD-17(8): 645–654, August 1998. 28. P. Feldmann and R. W. Freund, Reduced-order modeling of large passive linear circuits by means of the SyPVL algorithm, Proceedings of the IEEE/ACM International Conference on Computer-Aided Design, pp. 280–287, San Jose, CA, November 1996. 29. J. Xiong and L. He, Full-chip routing optimization with RLC crosstalk budgeting, IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 23(3): 366–377, March 2004. 30. Y. I. Ismail, E. G. Friedman, and J. L. Neves, Repeater insertion in tree structured inductive intercon- nect, Proceedings of the ACM/IEEE International Conference on Computer-Aided Design, pp. 420–424, San Jose, CA, No vember 1999. 31. H. Hu, D. Blaauw, V. Zolotov, K. Gala, M. Zhao, R. Panda, and S. Sapatnekar, Fast on-chip inductance simulation using a precorrected-FFT method, IEEE Tr ansactions on Computer-Aided Design of Integrated Circuits and Systems (T-CAD), 22(1): 49–66, January 2003. 32. H. Hu and S. S. Sapatnekar , Efficient inductance extraction using circuit-aware techniques, IEEE Transactions on VLSI Systems, 10(6): 746–761, December 2002. 33. G. Zhong, H. Wang, C. -K. Koh, and K . Roy, A twisted bundle layout structure for minimizing induc- tive coupling noise, Proceedings of the IEEE/ACM International Conference on Computer-Aided Design, pp. 406–411, Los Angeles, CA, June 2000. 34. M. A. El-Moursy and E. G. Friedman, Exponentially t apered H-tree clock distribution networks, IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 13(8): 971–975, August 2005. 35. M. A. El-Moursy and E. G. Friedman, Shielding effect of on-chip interconnect inductance, IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 13(3): 396–400, March 2005. 36. V. Mezhiba and E. G. Friedman, Inductive properties of high-performance po wer distribution grids, IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 10(6): 762–776, December 2002. Alpert/Handbook of Algorithms for Physical Design Automation AU7242_C042 Finals Page 881 23-9-2008 #2 42 Clock Network Design: Basics Chris Chu and Min Pan CONTENTS 42.1 Metricsfor Clock Network Design 882 42.1.1 Skew 882 42.1.2 Transition Time 882 42.1.3 Phase Delay 883 42.1.4 Area 883 42.1.5 Power 883 42.1.6 Skew Sensitivity to Process Variations 883 42.2 Clock Networks with Tree Structures 883 42.2.1 Method of Means and Medians 884 42.2.2 Geometric Matching Algorithm 884 42.2.3 ExactZero-SkewAlgorithm 885 42.2.4 Deferred Merge Embedding 887 42.2.5 Wire Width and Buffer Considerations in Clock Tree 888 42.3 Clock Networks with Nontree Structures 889 42.3.1 Grid 889 42.3.2 Spine 889 42.3.3 Hybrid 890 42.4 Clock Skew Scheduling 891 42.5 Handling Variability 893 References 894 A vast majority of VLSI chips are based on a synchronous sequential circuit design methodology. For these circuits, a clock signal is used to synchronize the operations of different components across the chip. Typically, this signal is produced by a clock generator circuit based on an external reference, and it is d istributed inside the chip by a clock network. Because the timing of the entire chip is controlled by the clock signal, a poor design of the clock distribution network will limit the performance of the chip. As the clock network connects the clock generator to a huge number of clocked elements (including latches, flip-flops, memories, and dynamic gates) all over the chip and it has high switching activity, it usually consumes a significant portion of the overall routing resources and of the total chip power. Hence the clock network must be carefully designed to optimize the performance of the chip, routing resource usage, and power consumption. This chapter discusses some basic issues in clock network design. In Section 42.1, the metrics used in designing clock networks are introduced. In Sections 42.2 and 42.3, algorithms to generate clock networkswith tree structuresandnontree structuresare described,respectively. In Section 42.4, the clock skew scheduling technique, which makes use of intentional clock skew to optimize per- formance, is introduced. In Section 42.5, clock network design techniques that focus on handling variability are presented. In Chapter 43, the clock network designs of several high-performance 881 . Alpert /Handbook of Algorithms for Physical Design Automation AU7242_C041 Finals Page 872 24-9-2008 #9 872 Handbook of Algorithms for Physical Design Automation 41.4.1 EFFECTS OF INDUCTANCE. magnitude of the excess repeater area when Alpert /Handbook of Algorithms for Physical Design Automation AU7242_C041 Finals Page 878 24-9-2008 #15 878 Handbook of Algorithms for Physical Design Automation TABLE. of Circuits, Systems, and Signal Processing, 13(6): 733–758, May 1994. Alpert /Handbook of Algorithms for Physical Design Automation AU7242_C041 Finals Page 880 24-9-2008 #17 880 Handbook of Algorithms