DesignCon 2005 Hybrid Stripline Analysis II: Propagation Characteristics, Crosstalk Effects and PCB Routability James R. Broomall, W. L. Gore & Associates, Inc. Email: jbroomall@wlgore.com Tamera A. Yost, W. L. Gore & Associates, Inc. Glen Walther, W. L. Gore & Associates, Inc. Gregg Wildes, Smiths Aerospace Abstract This paper is a follow-up to a study published at DesignCon 2004 which addressed the benefits of hybrid (mixed dielectric) stripline constructions by comparing the near-end crosstalk and insertion loss performance of hybrid stripline constructions to industry standard, homogeneous PCB stackups. The present study extends the analysis by comparing dispersion, far and near-end crosstalk, routing benefits and insertion loss performance of hybrid stripline constructions to industry standard, homogeneous PCB stackups. Data is obtained by direct measurement of test PCB traces and from simulated results. Finished PCB cost-performance considerations are also presented for the constructions evaluated in this study. Authors’Biographies James R. Broomall is a signal integrity engineer with W. L. Gore & Associates, Inc. in Newark Delaware. Jim has over 20 years experience in microwave and digital signal interconnects. He earned both M.S. and Ph.D. in physics from the University of Delaware and a B.A. in Physics from Franklin and Marshall College. Tamera A. Yost serves as Signal Integrity Team Leader at W. L. Gore & Associates, Inc., having 20 years of experience in the field of electrical and electronic engineering encompassing research and product development activities. She earned her M.S. and Ph.D. in Electrical Engineering at Drexel University and a B.S. in Electrical Engineering at the State University of New York. Glen Walther is the technical leader for Gore¦s PWB R&D Lab in Elkton, Maryland. Glen has been serving the PWB industry since 1989. Glen holds a B.S. in Industrial Arts & Education from North Texas State University, and a B.S. in Aerospace Engineering from the University of Texas at Austin. Gregg Wildes was the advanced dielectrics product specialist for the electronic products division of W. L. Gore & Associates, Inc. responsible for new application development. Gregg is currently with Smiths Aerospace and holds a Ph.D. in Materials Science from the University of Texas at Austin. Introduction Design constraints for today’s high data rate boards are forcing increased layer count when using conventional FR4 materials or the use of more expensive high performance dielectric materials. To mitigate these effects, there is increasing value in design topologies that allow more dense spacing of signal lines. A previous presentation1 indicated that a hybrid stripline construction provided lower attenuation and reduced near end crosstalk for a given stripline spacing than conventional constructions that use the same or similar materials for both core and prepreg. The hybrid construction uses a low cost core such as FR4 with a higher performance prepreg like GORE™ SPEEDBOARD® C Prepreg (SBC) to provide lower attenuation. Some questions remained concerning other measures of performance, specifically far end crosstalk, frequency dependence of time delay and differential signal performance. To address these concerns the hybrid construction is compared to the industry standard all FR4 stripline construction. Performance Considerations The effects of near end (NEXT) and far end (FEXT) crosstalk between coupled transmission lines can be thought of in various ways. A straightforward approach is to consider two lines as a differential pair, then look at the even and odd modes of propagation. In this approach, a step stimulus of the aggressor line is seen as a combination of one half even and one half odd mode, as depicted in Figure 1. This will result in a net zero input stimulus on the victim line. The response of the victim line is then the resulting superposition of the two modes. NEXT For NEXT, the characteristic impedance matrix of the coupled pair can be used to evaluate the reflection of the stimulus from both lines. For the case of a balanced pair, consider the following example: 60 10 Zc = 10 60 The odd mode impedance will be Zodd=Z11-Z12=50 Ohms and the even mode impedance will be Zeven=Z11+Z12=70 Ohms. The NEXT is then dependent on the even and odd mode impedance of the transmission lines feeding and sensing the aggressor and crosstalk signals. Assuming that the source lines are uncoupled lines of Zo=50 Ohm characteristic impedance, the odd and even mode impedances of the lines will both be 50 Ohms. In this somewhat simplified case, there will not be a reflection of the odd mode component, but there will be a reflection of the even mode component. Consider Figure 1 to help visualize the situation. Figure 1. NEXT explained as the response to a step aggressor signal resolved into even and odd mode components for the case of no odd mode impedance mismatch. If a 1 Volt aggressor signal having ½ volt odd and ½ volt even components is incident on one line, the odd mode will see no impedance mismatch, but the even mode will have a reflection coefficient of Z − Z even 50 − 70 ρ= o = = 0.167 Z o + Z even 50 + 70 A reflection of ρ times ½ Volt, or 83 mV, is seen on both lines. On the aggressor line, this is interpreted as an impedance mismatch, equivalent to a line impedance of 59.1 Ohms, calculated from this reflection relative to a 1 Volt stimulus. On the victim line, this is interpreted as NEXT of 8.3% relative to the 1 Volt stimulus. The impedance matrix is easily found using a 2D field solver for the geometry in question. Alternatively, it may be measured from the crosstalk and input impedance of one of the lines and reversing the calculation above. If the odd mode impedance of the lines is not exactly that of the source lines, both even and odd mode reflections need to be considered. A useful rule of thumb is that if the diagonal elements of Zc are 50 Ohms, the off diagonal impedance will be the crosstalk percentage. For other values of the diagonal elements, the percent crosstalk will not be the off diagonal value, but for practical cases, it will be close. As the lines are moved further apart, the off diagonal (coupling or mutual) impedances and the resulting NEXT will decrease. The discussion above assumes that the risetime of the step is shorter than the period of observation. In digital signals, the risetime of aggressor input signals is typically 25 to 50 percent of the bit width. The crosstalk voltage therefore has time to achieve its full amplitude. This is not the same situation as connector crosstalk, where the rate of change of current in rising and falling edges is significant. FEXT Far end crosstalk is inherently more complicated since it involves additional propagation characteristics of the transmission lines. But FEXT can also be analyzed by considering the aggressor to be separated into odd and even components. For simplicity, assume that equal amplitudes of the odd and even modes are launched into the coupled pair. If the even and odd modes propagate with the same velocity and attenuation, they will cancel each other as they exit the far end of the victim line, apart from any modal impedance mismatch reflections that occur there. The equivalence of modal propagation will rarely be the case however. Even when the stripline is centered exactly between the two ground planes, the stripline will likely be constructed from a core and prepreg that do not have exactly the same dielectric properties, resulting in different modal velocity and attenuation. Figure 2. FEXT explained as the response to a step aggressor signal resolved into even and odd mode components for the case when the even mode propagates faster than the odd mode. To visualize this effect, consider the case shown in Figure 2. If the attenuation is equal for both modes, but there is a different transit time for the two modes, one mode will arrive first. The leading edge of that mode’s transition will appear on the victim line before the other mode arrives to cancel its effect. The resulting voltage blip has a characteristic pulse-like shape, with a more rapid rise than fall time as seen in Figure 3. The amplitude and duration of the blip will be dependent on the length of the line since a longer line allows more separation between the modes due to their different speeds of propagation. Longer lengths allow the first mode to achieve more amplitude before the second mode arrives. If the modal velocities are the same, but the modal attenuation is different, the modes will not be able to cancel each other completely. In many cases, a combination of the two effects can be observed. The FEXT will tend to increase with longer length, but attenuation will tend to counteract this effect by decreasing the amplitude of the competing modes. The other effect seen in Figure 3 is the decrease of amplitude with increasing trace separation. This can be explained as a decrease of coupling between the lines, which causes the two modes to approach the same speed and attenuation. Figure 3. FEXT waveforms explained by different even and odd mode propagation velocities. Increased separation causes the modal velocities to approach each other. Other modes When the dielectric is not homogenous throughout the cross section, it has been recognized that additional modes of propagation can be stimulated. These have been studied in some depth theoretically2,3, but there are few experimental results in the literature. A complete discussion of this phenomenon is very much beyond the scope of this presentation, but some general comments are appropriate. There are two general classes of modes in addition to the expected modes that are “bound”to the striplines. One type results in leakage of the signal from the stripline into one or more parallel plate waveguide modes. These modes rob signal from the stripline, resulting in increased attenuation. The other very general type results in propagation in the direction of the stripline but typically having different propagation speed and attenuation than the bound mode. The generation of both of these additional modal types is dependent on transmission line geometry and dielectric homogeneity and consequently has propagation characteristics that are dependent on frequency. The description of FEXT above assumed even and odd modes propagating on the victim line that begin out of phase but with similar amplitude and propagate with different speeds and attenuation. In the case of these additional modes, they might typically launch with similar phase to that of the normal bound modes and by virtue of different propagation speed arrive at the end of the transmission line with different phase. The resulting phase difference will cause destructive interference that is a function of frequency and line length. This effect looks like a broad resonance-like band-stop filtering of the signal in the frequency domain and causes a variety of time domain effects on both the transmitted and FEXT signals. Some modes propagate away from the source and may cause FEXT that does not decrease rapidly with line separation. An example of this type of crosstalk is shown in Figure 4. Figure 4. FEXT caused in part by propagation of additional modes, interfering with the normal bound modes. Evaluation Boards In order to address these questions, a series of printed circuit boards were fabricated to explore the effects of asymmetry of placement of the stripline between the two ground planes and the use of homogeneous versus hybrid materials. Figure 5 shows schematically the stackup geometry and Table 1 lists the parameters as measured from cross-sections of the fabricated boards. Figure 5. Stackup parameters Table 1. Measured parameter values (mils) and Impedance (Ohms) Stackup Description a b W t Z0 1 Balanced Homogeneous 9.3 11.2 7.55 0.36 59.2 2 Unbalanced Homogeneous 13.9 9.6 7.69 0.36 55.8 3 Balanced Hybrid 10.1 7.1 7.51 0.36 56.7 4 Unbalanced Hybrid 13.9 5.6 7.55 0.36 54.6 The stackups were chosen to demonstrate the effects of material homogeneity and unequal thickness layers of dielectric surrounding the stripline. Stackup 1 was a homogeneous design of all FR4(Nelco 4000-6) and was used as an industry standard baseline construction. Stackup 2 was also a homogeneous design of all FR4 but with the stripline trace offset with respect to the top and bottom ground planes. Stackup 3 was a hybrid construction of FR4 and SBC intended to provide the same capacitance per unit length between the trace and each ground plane, therefore demonstrating a “balanced”configuration. And lastly, stackup 4 was intentionally designed to emphasize the effects of an offset in a hybrid construction of FR4 and SBC by promoting stronger coupling of the signal through the SBC. Microphotographs of the stackups are shown in Figure 6. The trace thickness of 0.36 mils is less than was originally planned, causing the trace impedance to be higher than expected. This causes a somewhat higher trace resistance, but the higher impedance lowers the conductor loss while having little effect on the dielectric component of the loss. Figure 6. Microphotographs of the 4 stackups. The left two have core and prepreg of FR4, while the two on the right are FR4/SBC hybrids. The top two stackups are more balanced than those on the bottom. The traces are slightly narrower than the nominal 8 mil width. When spacing is described in this presentation, the nominal width is assumed. For example, an 8 mil spacing results from a 16 mil pitch. As a result of the higher single line impedance, the differential impedance at 8 mil spacing is close to 100 Ohms. The board layouts all included coupled lines at 8, 16 and 20 mil spacing centered between via fences 1 inch apart. The 20 mil diameter plated through holes were spaced at a 0.25 inch pitch parallel to the stripline traces. The 1 inch spacing was intended to not unduly influence the propagation characteristics while maintaining reasonable contact between the ground planes. The stripline traces were electrically fed by 3.5mm or 2.9mm edge launch coaxial connectors. The launches from the coaxial connector pass through a very short section of co-planar waveguide to a via down to the stripline. A 0.23 inch length of stripline leads to the coupled section of striplines being investigated. Typical return loss of –20 dB was achieved though 20 GHz. Most data was taken on 13 inch long sections of coupled stripline. Measurements in both time domain and frequency domain are needed to evaluate the transmission properties of interest. Response to a step voltage source, eye diagrams using a pattern generator and Sparameters from a 4-port vector network analyzer were used to acquire the needed measurements. Experimental Results and Discussion Single-Ended Measurements Measurements were taken for near end (NEXT) and far end (FEXT) crosstalk on lines at 8, 16 and 24 mil spacings using four ports on a time domain sampling oscilloscope. Short lengths of low loss 50 Ohm coaxial cable were used to connect the scope ports to the SMA compatible board launch connectors. The signal paths to and from the board were carefully deskewed. A TDR source of one of the scope heads was used to provide a step aggressor signal to one of the launch connectors. The aggressor step had an amplitude of 250 mV and a 10-90% risetime of 30 ps as measured on a path through the connecting cables using a short SMA bullet. The TDR step was chosen as an aggressor signal since it provided a clear response for NEXT and FEXT. The 30 ps risetime is appropriate for digital signals in the 10 Gbps range. Input impedance, NEXT, FEXT and transmitted signal amplitude could be viewed with a single configuration. Sample data is shown in Figure 7. The top signal trace is the NEXT response, showing the typical step-like form of transmission line NEXT. The second signal trace shows a typical FEXT blip on the victim line. The third signal trace is the input impedance, displayed in units of Ohms at 5 Ohms per division. The bottom trace is the signal exiting the aggressor line. All signal traces except for the impedance response are displayed at 10 mV per division. All of the tested boards showed typical NEXT step responses like that of the sample with varying amplitudes depending on stackup and trace separation. Figure 7. Sample display showing NEXT, FEXT, input impedance and transmitted signal amplitude. The NEXT results for the four board stackups are summarized in Figure 8. There is not much difference between the balanced and unbalanced stackups for a given material set, but there is a significant difference between the hybrid and homogeneous costructions. The hybrid boards show approximately a 30% improvement in NEXT over the homogeneous boards. Figure 8. Near end crosstalk as a function of trace spacing for the four board stackups. The far end crosstalk results are shown in Figure 9. The unbalanced Hybrid board shows substantially more crosstalk than the other stackups at all spacings. The crosstalk waveforms for this board were used to assemble Figure 4. By contrast, the crosstalk waveforms for the balanced homogeneous board illustrate the expected falloff with spacing as shown in Figure 3. The balanced boards of both material sets showed very similar results at 8 mil and 16 mil spacings. The unbalanced Homogeneous material board showed very low FEXT at all spacings, although there was some evidence of the behavior seen in Figure 4, although at a small amplitude. Figure 9. Far end crosstalk as a function of trace spacing for the four board stackups The homogeneous results are somewhat surprising at first glance. If the material is truly homogeneous, the even and odd modes should propagate at the same speed, producing very little FEXT. This should be true regardless of trace centering. Inspection of the material specifications for different thickness of core and prepreg4 shows a fairly broad range of possible dielectric constants depending on thickness. This may be the root of the difference in these two examples. The data from the hybrid boards both showed evidence of the ringing waveforms illustrated in Figure 4. This is likely due to the generation of the additional waveguide modes described earlier. The unbalanced example generates significantly more FEXT than the other samples. On the other hand, the more balanced example showed FEXT comparable to the balanced homogeneous sample at the lesser spacings that are more attractive from a board real estate perspective. Eye diagram measurements were taken on the single trace paths. With the obvious exceptions, those results are very similar to those presented later for differential paths, since the attenuation is very similar. They will not be shown here. There is no obvious way to isolate measurement of the additional modes caused by lack of homogeneity. There is however the ability to look at their effect in the frequency domain. The lines at 8 mil spacing were tested using a 4-port vector network analyzer. The phase data from the through path of one of the lines for each board was unwrapped to provide phase as a function of frequency. This data was inspected to be certain there were no regions where the phase data was not monotonically increasing due to frequency step size. The phase data was then divided by frequency at each point and converted to time delay. The estimated time delay of the launches was subtracted and the result scaled by length to time delay in ps/in. Figure 10 is a graph of these results plotted versus frequency for each board. Note that all responses show the typical delay increase at low frequency caused by an increase in the conductor inductance expected at low frequency. Figure 10. Time delay versus frequency extracted from S-parameter measurements. The time delay results show the expected difference in effective dielectric constant for the hybrid versus the homogeneous stackups. The homogeneous stackups have measurably different delays. The likely explanation is differences in dielectric constant of the FR4 cores and prepregs. This possibility is consistent with the FEXT results presented earlier. The unbalanced hybrid results are striking. Starting as low as 6 GHz there appears to be a transition to a different mode of propagation, or possibly combined modes. The balanced hybrid remains well behaved to above 12 GHz. Compare these results to the insertion loss plots in Figure 11. All but the balanced homogeneous construction show some evidence of irregularity at higher frequency. The dip in S21 of the unbalanced hybrid construction coincides with the delay anomaly seen in Figure 10. Figure 11. Single-ended through path insertion loss for board stackups. Differential Measurements In addition to the single-ended results, S-parameter and eye diagram measurements were taken on the differential paths created by the coupled traces at 8 mil spacing. The differential insertion loss is shown in Figure 12. The balanced constructions show very nice performance. Both unbalanced versions show some anomalies, but they are not as severe below 15 GHz. Figure 12. Differential through path insertion loss for board stackups. Eye diagram measurements were taken with a 1 Volt differential source at 10 Gbps using a 27-1 PRBS pattern. Figure 13 shows the source and board path results. Figure 13. Differential eye diagrams for input and 13-inch traces for each stackup. The displayed eye diagrams were measured with 125 mV and 20 ps per division settings. The increased high frequency insertion loss of the all FR4 versus FR4/SBC construction is evident in the reduced eye opening height as shown in Figure 14. The associated increase in inter-symbol interference is the major contributor to deterministic jitter in passive components. Peak-to-peak jitter measurements of the eye diagrams using the oscilloscope are summarized in Figure 15. The measurements include the contribution of 10.8 ps of peak-to-peak jitter from the pattern generator. The hybrid construction, regardless of stackup, shows a distinct jitter advantage versus an all FR4 construction as seen in the eye diagrams. Figure 14. Eye diagram maximum eye height for all stackups. Figure 15. Eye diagram peak-to-peak jitter for all stackups. Unfortunately, multiple adjacent lines were not included in the these boards to allow differential crosstalk measurements to be made. Conclusions The study described in this paper solidifies the notion that understanding the propagation characteristics of complex structures is difficult. Various modes of propagation exist in most cases that can cause signal degradation. This is true even for the all FR4 balanced stripline construction used widely in industry today. As expected the hybrid construction of FR4 and low loss GORE™ SPEEDBOARD® C Prepreg provides a benefit in signal transmission due to lower attenuation consistent with a lower effective dielectric loss tangent. For a given trace spacing, the near end crosstalk for the hybrid construction is noticeably less due to weaker coupling between the traces and stronger coupling to the ground plane. For a balanced construction, the far end crosstalk of the hybrid and homogeneous dielectric stackups were comparable. If particular care is used in matching the core and prepreg materials for the homogeneous construction, the far end crosstalk may be minimized. It is believed that there also exists an optimum thickness ratio for the hybrid construction that may minimize the far end crosstalk. While near end crosstalk is dependent upon the impedance matrix values, far end crosstalk is a function of the modes of propagation. The causes of far end crosstalk are complex and are affected by the launch, trace length, and interaction with other planar features on the board such as grounding vias. Some initial investigations have shown that placing grounding vias close to the trace can minimize the generation of unwanted modes of propagation. The literature contains a variety of suggestions for inhibiting these modes5,6,7. When compared to an all FR4 construction, the board design benefits of this work include the use of narrower lines, closer spacing, and reduced layer thickness. In addition, the performance of a hybrid construction is significantly better than all FR4 and comparable to a homogeneous mid-range performance dielectric material set. Various cost analyses have been conducted that repeatedly show cost savings of 20% using a high performance prepreg with a low-cost FR4 over a mid-range costperformance material in a balanced stripline construction. The reduced cost originates from the material cost benefits of continuing to use low-cost FR4, reduced board layer count, and reduced manufacturing costs associated with lower layer count boards. To complete the study of performance of hybrid stripline constructions, some additional work needs to be done. We have shown that hybrid stripline structures offer a variety of benefits. To understand the complex length-dependent modes of propagation a detailed model is needed to allow for optimization of the design parameters. Specifically, the literature suggests that scaling of layer thickness can shift the interaction of the unwanted modes to a frequency above our region of interest for 10 Gbps. If this is true, the use of hybrid structures for narrow traces in dense designs would be an obvious choice. One of the promising aspects of this study seems to lie in the use of a hybrid stripline structure for differential signals where the effects of unwanted modes are greatly reduced. The data in this study showed a 25% reduction in insertion loss at 20 GHz for the hybrid stackup. A further study of crosstalk between differential pairs as a function of pair-pair spacing is planned. References 1. N. Hudson, T. Yost and G. Wildes, “Potential benefits of mixed dielectric stripline”, DesignCon 2004, Feb. 2004. 2. W.L. Langston, J.T. Williams, D.R. Jackson and F. Mesa, “Spurious radiation from a practical source on a covered microstrip line”, IEEE Trans. Microwave Theory Tech., vol. 49, pp. 2216-2226, Dec. 2001. 3. M. Tsuji and H. Shigesawa. "Behavioral feature of fast-wave modes on printed-circuit transmission lines of open and packaged types." 2001 MTT-S International Microwave Symposium Digest, pp. 863-866 vol.2, 2001. 4. “N4000-2, -6 and -6 FC Dielectric Properties Table”, pp1-2, Park Nelco, 2003 5. T. Tischler, M. Rudolph, A. Kilk and W. Heinrich. "Via arrays for grounding in multilayer packaging - frequency limits and design rules." 2003 MTT-S International Microwave Symposium Digest, pp: 1147-1150 vol.2., 2003. 6. T. Yuasha, T. Nishino and H. Oh-hashi, “Simple design formula for parallel plate mode suppression by ground via-holes, ." 2003 MTT-S International Microwave Symposium Digest, pp. 641-644, 2004. 7. K-P. Ma, J. Kim, F-R Yang; Y Qian, T. Itoh, “Leakage suppression in stripline circuits using a 2-D photonic bandgap lattice”, 1999 MTT-S International Microwave Symposium Digest, pp.:73 –76, Jun. 1999.