DesignCon Loaded Parallel Stub Common Mode Filter. Predrag Acimovic, PMC-Sierra, Inc

Transcription

1 DesignCon 2008 Loaded Parallel Stub Common Mode Filter Predrag Acimovic, PMC-Sierra, Inc

2 Abstract EMI radiation problems are usually due to certain unwanted common mode signals. This paper presents a novel structure that can be used to filter selected frequencies of the common mode present on high-speed differential signals, therefore preventing these signals from reaching effective radiators and reducing the chance for EMI problems. The structure is very simple and is suitable for implementation on printed circuit boards or backplanes, but it can also be used within the chip, either on die or package substrate. While the structure affects low frequencies, it can be also used as an equalizer. Biography Predrag Acimovic is a Technical Advisor in PMC-Sierra s Mixed-Signal Design Group. His responsibilities include analysis and design of packages for signal and power integrity, with a focus on the design of passive RF components on chip. He joined PMC- Sierra in 1996 as an applications engineer, where he supported customers using PMC- Sierra products. Prior to joining PMC-Sierra, Mr. Acimovic held design-engineering positions at Microtel Pacific Research Ltd. s Radio and Satellite Division, designing microwave active and passive MIC/MMIC circuits; and at EI s High Frequency Division Microwave Radio R&D lab. Mr. Acimovic received a Master of Engineering degree from Simon Fraser University in 1990 and a Diploma of Electrical Engineering from the University of Belgrade in 1978.

3 Introduction Components of electrical circuits radiate electromagnetic waves into the space around them. The circuits that operate at higher frequencies are more likely to effectively radiate into the surrounding environment. High-speed SERDES circuits, especially their interconnections, are therefore a likely candidate to cause radiation problems. Spectrum use is restricted via regulations issued by commissions delegated by governments. Some devices claim the spectrum as part of their intended operation, but other devices, like computers and displays, radiate unwanted emissions that can, in some cases, compromise the operation of other electronic devices. For that purpose, the United States Federal Communications Commission (FCC) and Europe s CISPR regulate the allowable amounts of radiated emissions, depending on the place of use. Class A targets an industrial environment and class B targets a residential environment. Class A specifications are a bit more relaxed than class B. EMI requirements cannot be waived. A product has to pass the test in order to be sold. The problem of meeting the EMI specification is not an easy one to solve because there is no reliable tool to predict the amount of EMI emissions accurately. It is important to be able to control the problem effectively and cheaply, without too much shielding. In some cases, like in the case of open-box equipment, shielding is not at option at all. In equipment that uses a high-speed SERDES devices, most of the radiation is caused by the common mode propagating along the PCB traces and reaching efficient radiators, like connectors or cables attached to equipment. Here we present a band-stop filter that will filter the common mode and prevent it from reaching the efficient radiators. Distortions that generate common mode signals There are several distortions of high-speed differential signals that will generate the common mode, but not all common mode signals produce dangerously high levels of EMI. It is usually the discrete frequencies that are most damaging. The usual sources of the common mode signal are: a) Uneven rise and fall times of the two differential outputs; b) Duty cycle distortion that results in delayed rise or fall edges with respect to opposite kind of edge; c) Skew in output signals; d) The two output signals not equal in amplitude; and e) One of the output signals can be filtered more than the other. The linear generation of common mode noise produces a common mode signal with spectrum that is distributed. Therefore, this kind of signal distortions less likely to produce levels EMI that fail to meet requirements. Linear types of common mode generation are cases c), d) and e). The EMI measurements are performed with a spectrum analyzer that has a specific resolution bandwidth, so the distributed power spectrum of the common mode signal is not likely to cause enough problems to not pass the test. Problems arise in the case of non-linear kind of common mode signal generation. These are the cases a) and b).

4 Nonlinear common mode generation a) Uneven RT and FT b) DCD Distortion affects alternatively N (blue) and than P (red). Linear common mode distortions affect only one in the pair of signals that form a differential pair. Non-linear common mode distortions alternatively affect one then the other signal of a differential pair, as can be seen in the above picture. It acts as a strong second order non-linearity and produces discrete spurs at the harmonics of the data rate. Non-linear common mode generation produces a common mode signal that does not depend on the data pattern, but rather on the probability of whether there is a transition or not. In paper [6] the expression for the single side power spectral density was derived. SSPSD 2 R A 4 ( )&( ) [ ( / 2)] a b = 2 Sinc ω R k= 0 [ δ ( f k T )] Equation 1 In this expression, R is rise/fall time and A is amplitude of the common mode waveform. T is symbol interval. The spurs will occur on DC and all multiples of the symbol rate. It is important to note that the R (rise/fall time) of the common mode signal should be much smaller than T, as the Sinc function is practically constant within the first harmonic of symbol rate if R is much smaller than T (R< 10% of T). Therefore, the amplitude of spurs should depend mainly on the product of R and A. The product of R times A is equal to the energy of one triangle-shaped pulse, so it is not the value A, or peak of the common mode signal, that is important as much it is the energy of the common mode pulse. Spur power also depends on the fourth power of Sinc function. For that reason, if the rise/fall time R is a significant portion of the symbol interval T, then a reduction in the spur amplitude should occur. Sinc term in power spectral density shows that, for cases where the slew rate is reduced such that it becomes a significant portion of the symbol interval, we can achieve a reduction of spur. To illustrate the effect of non-linear vs. linear common mode distortion, here are plots of the spectrums of both differential and common mode signals.

5 Figure 1 Spectrum of differential signal (blue) and common mode (red) in case of nonlinear common mode distortion (difference between rise and fall times 1/16th of UI) Figure 2 Spectrum of differential (blue) and common mode (red) signals in case of linear common mode distortion (skew between p and n signals 1/16th of UI) Spectral plots in Figures 1 2 show the distinct difference between linear and non-linear common mode distortion. Abscissa is normalized to symbol rate. Linear distortion does not produce new frequency components that don t already exist in the differential signal. This is because linear distortion that causes the common mode affects only one of the signals that constitutes the differential signal, therefore it contains similar spectrum to the differential signal, except for some filtering. In the example presented in Figure 2, the filter is high pass. This type of distortion is associated with packages, connectors, cables, PCB traces, and vias. Non-linear distortion that creates the common mode can only be generated in active devices. This type of distortion acts as strong even-order non-linearity and produces spectrum that is entirely different than the differential spectrum shown on Figure1. The distinct discrete spurs appear at multiples of the symbol rate and are the main problem during EMI testing.

6 Figure 3 FFT of the differential (blue) and common (red) mode signals in case of pattern 1010 with perfectly balanced traces Figure 4 FFT of the differential (blue) and common (red) mode signals in case of pattern 1010 with output unbalanced with 200fF The non-linear distortion that produces the common mode followed by the linear distortion does not increase the level of spurious components at symbol rate and multiples of symbol rate. In Figure 3, there is a spur at 6GHz and there are lower spurs at higher multiples of 6GHz. Figure 3 shows the FFT of the differential and common mode signals at the output of the driver that is perfectly balanced. Figure 4 show the FFT of the differential and common mode signals of the driver that feeds the package and PCB that is not perfectly balanced but has approximately 200fF extra on one of the signals that form the differential pair. The plots above are using an alternating 1 and 0 pattern in order to show how the linear and non-linear common mode distortions generate spectrum. The alternating 1 and 0 pattern is deterministic and the differential signal has strong component at half a symbol rate 3GHz. Linear common mode generation does not produce new frequency components, but as there is already a strong 3GHz due to

7 differential signal, it will produce a common mode signal with a 3GHz spur. If the data pattern were random, there would not be a strong 3GHz differential signal or a 3GHz common mode spur. Spurs will only arise at 6GHz and multiples of 6GHz. The level of a 6GHz spur did not increase due to introduced skew in p and n signals. Basically, common mode resulting from skew introduced in cables when the data pattern is random is not responsible for radiation and EMI. Unfortunately, if there was non-linear distortion that generated non-linear common mode at the input of the cable, then the cable, acting as a good antenna, will radiate and cause EMI problems. Narrowband Band-stop Common Mode Filter One of the methods to reduce common-mode-based problems is to filter the common mode, but it is important not to distort the differential signal. Use of common mode chokes is one way to achieve both goals. Ferrite-based chokes need to be mounted on top of the PCB, requiring PCB vias, which can be detrimental to performance at high frequencies. If the differential lines are on top of the PCB, then they can radiate. Also, the mounting pads for the common mode chokes will present discontinuities. So, at frequencies of 8Gbit/s and higher, use of common mode chokes may not be preferable. The basis for and idea behind the narrow-band common mode filter can be shown using the S-parameters and power flow graphs. A differential signal is sent using a coupled transmission line. We can assign ports to this structure as shown in Figure 5. port1 port2 port3 port4 Figure 5 Port assignment of a pair of coupled transmission lines The S-parameters can be used in calculating the transfer function of the pair of coupled transmission lines. We can calculate both differential and common mode transfer functions. If we draw the power flow, we can write the expressions for both transfer functions. In the case of the differential mode, we can assume that the differential signal is applied to port 1 and port 3, which means that the voltages on these two ports are equal and opposite in sign. In the case of the common mode, both ports 1 and 3 are excited with identical signal. 1 S21differ = ( S21+ S43 S41 S23) 2 1 S21comm = ( S21+ S43 + S41+ S23) 2 (1) (2)

8 S21differ is a transfer function of differential mode and S21comm is transfer function of common mode. From the second equation (2), it is obvious that we can reduce S21comm, therefore attenuate the common mode, if we split the signal from port 1 in two equal signals and invert one of these two signals, doing the same with the signal from port 3. So, if S21=- S41 and S43=-S23, the S21comm=0. At the same time, S21diff is maximized, assuming the transmission lines are symmetrical (S21=S43). This conclusion about how we can attenuate the common mode without attenuating differential mode can be argued more simply without the S-parameters and flow graphs. Basically, the voltages on ports 1 and 3 are identical for the common mode. If terminals 1 and 3 are connected though the circuit that produces 180 degrees at certain frequency, then the common mode components from terminal 1 and 3 will be opposite in sign and same in amplitude, and they will neutralize each other. At the same time, the differential components will add in phase and will not be attenuated. The main difficulty is to split the signals perfectly and invert them over all frequency components of interest and then combine the signals without reflections. That can be a very difficult task. One option is to selectively invert some portions of the signal. As the most damaging frequencies resulting from common mode are harmonically related to the symbol rate, it should be sufficient to invert the symbol rate frequency and perhaps the second or third harmonic. For this, we can use a narrowband 180-degrees inverter. If we refer to the random digital signals spectral densities related to differential and common mode signals, the worst discrete frequency components of common mode signals fall in the frequency ranges where the differential signals spectrum is having least amount of power. At first glance, the best approach would be to filter only frequencies that need to be inverted (Figure 6), and pass them from one line to the other. This would work for steady-state but not for random data, where the spectral components change in time depending on the particular data sequence. The narrowband filters have large-phase shifts; therefore the delay through the filters will not make this possible to work as the common mode filters. Of course, the common mode discrete frequency components related to the symbol rate do not exactly depend on the data sequence, but they depend on the density of transitions. If the data transition density is more or less the same, even this contraption with narrowband filters should achieve relatively good common mode suppression and, at the same time, should not distort the differential signal too much, as long the narrowband filters present high impedance in frequency ranges where the differential signal has a significant amount of spectral components.

9 RS Z0m Z0m Vps R=Z0m Fr1 Fr2 Frn 180 deg 180 deg 180 deg RS Z0m Z0m Vns R=Z0m Figure 6 Possible implementation of common-mode filter In Figure 6, Fr1, Fr2 and Frn designate the frequencies that the narrowband-pass filter passes. The filter and 180-degree phase shifter need to have total phase shift of 180 degrees at frequencies Fr1, Fr2 and Frn respectively. A slightly different approach is presented in Figure 7. A differential transmission line is loaded with a parallel stub that is terminated at its end with resistance 2*Rpar. The length of the parallel stub is one-quarter wavelength at the frequency at which we wish to attenuate the common mode. It is best to separately analyze the differential and common modes. The circuit in Figure 7 performs differently in respect to differential and common mode excitation. RS Z0m Z0m Vps R=Z0m RS Z0m Z0m Vns R=Z0m Zos 2*Rpar length Figure 7 Parallel stub terminated in load 2*Rpar

10 For common mode excitation, the parallel stub inverts the common mode from one line so when it adds to the common mode signal from the other line the total common mode signal becomes zero. The reduction of the common mode can only be achieved over a narrow band, but because the most damaging common mode spurs are at frequencies related to symbol rate, the common mode spurs can be reduced very effectively. The differential mode will be distorted, as the parallel stub will present the discontinuity, but proper design of the parallel stub can result in distortion of the differential mode that is very small and, in some cases, we can even improve the differential signal quality. First, let us analyze the performance of the circuit from Figure 7 in the case of common mode excitation (Vps=Vns). The circuit is completely symmetrical with respect to both signal transmission lines, so the voltages along each transmission line will be identical. Also, voltages along the parallel stub will be identical. Therefore the voltages on both terminals of the resistor, with value 2*Rpar, will be identical, which means that there will be no current flowing through 2*Rpar. As there is no current flowing through 2*Rpar, the parallel stub looks like a transmission line with an open end. The value of Rpar should not factor in common mode transfer function. Basically, the quarter wavelength of parallel stub transforms the common mode open at its end to short at the junction between the differential transmission line and parallel stub. So, the common mode is effectively shorted to ground with the parallel stub. The response of a circuit to differential mode excitation Vps= -Vns is quite different. The circuit is again symmetrical with respect to differential mode excitation, but in this case, there is a short to ground along the line of symmetry, whereas in the case of common mode excitation, there was an open circuit along the axis of symmetry.

11 Z0m Z0m Zstub Z0m Z0m Rpar Ground along axis of symmetry Equivant cricuit used to calculate Differential-mode S-paramters Z0m Zstub Z0m Z0m Z0m Rpar Open along axis of symmetry Equivant cricuit used to calculate Common-mode S-paramters Figure 8 Differential-mode and Common-mode equivalent circuits First, we will derive the expression for S-parameters. Zstub is the impedance that loads the high-speed transmission line at the junction of the two. Rpar + j Zos tan( β stub lengthstub ) Zstub = Zos Zos + j Rpar tan( β stub lengthstub ) Zos is the characteristic impedance of each transmission line that form the parallel stub. The stub has length lenght stub and propagation constant β stub. The length of the stub should be such to produce 90 degrees at the frequency of the common mode spur. To calculate the differential and common mode S-parameters of this circuit, note that the sum of the incident and reflected wave immediately at the left of the junction have to be the same as the sum of the incident and reflected wave immediately at the right of the junction. Zstub Zom Zx = Zstub Zom = Zstub + Zom Zx Zom Zom S11 = = Zx + Zom 2 Zstub + Zom 2 Zstub S21 = 1+ S11 = 2 Zstub + Zom The above expressions for S-parameters of the parallel stub common mode filter are valid for both the common and differential mode of propagation. In the case of the common mode, the Rpar is effectively infinite, so the expressions can be simplified even more.

12 The common mode transfer function contains a pole frequency where equal to π/2. β length is stub stub j Zos Zstub = where Ω = tan(β stub lengthstub ) Ω 1 S11CM = 1 j 2 Zos 1 Zm Ω 1 S21CM = 1+ j Zom Ω 2 Zos The pole of the common mode S21 is primarily dependent on the length of the parallel stub, as the tan function has a much higher sensitivity around π/2 than the possible change of ratio between characteristic impedances of transmission line (Zom) and parallel stub (Zos). The common mode S21 dependency on the characteristic impedance values of the transmission lines is negligible and that Rpar value does not figure at all in the expression, so all these circuit parameters can be adjusted to achieve the best performance of the differential transfer without affecting the attenuation of the common mode spurious frequencies too much. A quick analysis shows that even a ratio of Zos/Zom>2 gives us enough margin to cover the FR-4 PCB manufacturing tolerances and dielectric material properties and environmental variations to achieve 10dB attenuation. Figure 9 shows the parametric sweep of the ratio Zos/Zom from 1 to 2 in 0.2 increments. The widest is the band-stop filter with Zos/Zom=1 and the narrowest is for Zos/Zom=2. Frequency is normalized, and it is shown only out to a normalized frequency of 2 but, as the tan function is periodic, the transfer function will repeat in frequency. This shows the benefit of using the transmission line to build the parallel stub, as one stub will filter not only the first harmonic but also all odd-numbered harmonics. Basically, to filter all odd and even harmonics of symbol rate, we would need only two parallel stubs. Later, we will discus how to design the common mode filter for both odd and even harmonics of symbol rate.

13 Figure 9 Common mode S21 of the parallel stub common mode filter Figure 9 shows that parallel stub vs. line characteristic impedance ratio Zos/Zom=1 will produce a band-stop of about 20% and Zos/Zom=1 will produce a band-stop of about 10% for attenuation of more than 10dB. Differential modes S21 and S11 depend on a number of parameters. The common mode attenuation is determined by the length of the parallel stub, so it is possible to independently chose the relationship between Zom, Zos and Rpar. Figure 10 shows the calculated values for differential modes S21 and S11 for Zom=50, which is the most common impedance of the transmission lines used for high-speed data communication. The values of Zos and Rpar are varied. The plots of differential mode S21 and S11 in Figure 10 are normalized to the symbol rate. On first inspection, it is obvious that S21 is exhibiting loss, which is not a desirable property. Closer inspection of the differential mode S21 of the parallel stub shows an interesting and potentially desirable property of S21 monotonically increasing in range from DC to the symbol rate for certain ratios of Zom, Zos and Rpar. For a value of Rpar smaller than the characteristic impedance of parallel stub Zos, S21 is increasing in range of DC to the symbol rate. This fortunate property of the differential mode of the parallel stub can be used to compensate for losses in the system. Although the effect of increasing differential S21 is only in range from DC to the symbol rate, this is the most important frequency range that contains the most of the digital signal power. Rpar=Zos produces the flat frequency response of differential S21. For set Rpar=Zps, the ratio of Zos/Zom determines the amount of attenuation of the differential S21. For Zos=Rpar=Zom, the amount of attenuation is 3.54 db. The higher the ratio Zos/Zom, the less attenuation results for differential S21. In case of Zos/Zom=2, the attenuation is slightly less than 2dB, and for Zos/Zom=4, the attenuation is only 1dB.

14 System transfer attenuation at Nyquist frequency, half of the symbol rate, is most responsible for the eye opening and bit error rate. If the system has attenuation at the Nyquist frequency of a certain amount, it is possible to find the desired combination of Zos and Rpar to achieve an almost flat frequency response from DC to Nyquist frequency. This kind of system will have no smaller eye opening than the system without the parallel stub, with the advantage of less jitter, and of course we should not forget about attenuated common mode. Another important parameter in system performance is the return-loss of every element that comprises the transmission system. The return-loss is also a parameter of the ratios of Zom, Zos and Rpar. The higher the Zos/Zom ratio, the better the return-loss. If we can achieve Zos/Zom=4, the return loss would be almost 20dB at frequencies near the Nyquist frequency. We also can note that the return-loss improves for higher frequencies, which also compensates for the usual tendency for return-loss to be worse at higher frequencies. Since the circuit eliminates common mode spurs, a small penalty in returnloss and a slight decrease in eye opening should not be a difficult compromise.

15 Figure 10 Differential mode S21 and S11 as a function of Zos/Zom and Rpar

16 Figure 10 sows that an increase of the Zos reduces the insertion loss. It improves the return-loss. The increase of Rpar to Zos value reduces the slope of the transfer function. It makes the transfer function more constant. Zos and Rpar can be used to compensate for certain insertion loss of the transmission line but the more insertion loss we need to compensate for, the more insertion loss needs to be created by the parallel stub circuit. One way to compensate for this is to implement the parallel stub circuit inside the IO driver. With this method, we should be able to compensate for the parallel stub losses while maintaining the peak-to-peak output swing at the same time. For low data rates, it may be difficult to implement the parallel sub circuit using transmission lines on the die. For higher data rates above 10Gbit/s, it may be possible to implement the transmission line parallel stub circuit on the die. Implementing the parallel stub on-die may be advantageous as we could vary the Rpar and Zos. One way to vary Zos is to implement the transmission line as LC circuit. RS Z0m Z0m Vps R=Z0m RS Z0m Z0m Vns R=Z0m Ls1 2*Rpar Cs1 Cs1 M1 Ls1 Figure 11 Parallel stub implementation using LC elements M1 is the mutual inductance. Mutual inductance has a different effect on the common mode and differential mode and it is a subject we will discuss after analyzing the circuit in Figure 12. We can approximate the LC circuit as a transmission line with characteristic impedance equal to the square root of the ratio of Ls1 and Cs1. The values of Ls and Cs1 determine the phase shift. From circuit theory, we know that the 90-degree phase shift is at frequency F90 given by the product of Ls1 and Cs1. Ls1 Zch LC = Cs1 Fr 90 deg = 2 π 1 Ls1 Cs1

17 So, this circuit which is an approximation of the parallel stub has more potential than the parallel stub as we can tune this circuit in such a way to change Ls1 and Cs1 and Rpar, and therefore change where the pole of the common mode transfer function is by setting Ls1 and Cs1 to produce a 90-degree shift in each leg of Figure 11. The limitation of implementing a parallel stub with a transmission line is that it is limited with the highest impedance that can be built. If we are using microstrip, the range of characteristic impedances that can be reliably produced is likely in the range of 25 Ω to 125 Ω. This impedance can be increased using coplanar strips but it is still limited and, once the characteristic impedance of the trace is chosen, it can not be changed. Using lumped element L and C, we can design the L/C ratio to be rather high, therefore making the differential transfer-function losses in the parallel stub very small, as well improving the return loss. Also, in case of LC implementation, it is possible to easily change C, and possibly even L, by changing its magnetic field or by switching in more or fewer inductors. Variable capacitance can be implemented in a number of ways as capacitance of any inversely biased p-n junction depends on the voltage across the junction. This means that we can have full programmability if the LC circuit is implemented. The best option is when the parallel stub is implemented in the chip, as then the source impedance can be adjusted for the best return-loss and the best transfer function. The parallel stub as part of the chip also makes it possible to adjust for the losses in the parallel stub junction, as the output of the chip will still have the same output level as it would if there were no parallel stub. The advantage of implementing this structure in-chip is that adjusting the common mode frequency as a feedback can be applied. The two output signals, p and n, can be combined to give the common mode signal. In fact, the tap in the middle of the 2*Rp resistor can be used. This common mode signal can be down-converted with the clock signal to give DC representation of the spur level. The filter is best adjusted for a DC voltage closest to zero. From the analysis of the parallel stub circuit, it follows that the higher the Zos, or ratio of Ls1 and Cs1, the differential mode will exhibit less insertion loss and better return-loss. This was illustrated in Figure 12, because the LC circuit 90-degree phase shift is important for the common mode, which is high Q. As it is open-ended effectively, the structure has a very nonlinear phase and rapid transition through the 90-degree operating point. The sensitivity to Ls1 or Cs1 value changes is very high. One effective way to minimize that is to use coupled inductors, as indicated in Figure 14. This is exactly what we need, as we need to increase the effective Zoc of the stub in order for the differential mode to decrease differential losses. In the case of the common mode, the magnetic fields oppose each other, so the effective inductance will decrease, which is exactly what we need to reduce the sensitivity to component value variation. Therefore, large coupling between the inductors is very beneficial. The only limit in achieving very large Ls1/Cs1 ratios is the parasitic capacitance; otherwise we would be able to achieve infinite Zoc for the differential mode.

18 One way to increase the ratio of Ls1/Cs1 and maintain Cs1 at accurately manufacturability values is to split the LC 90-degree phase shift in several LC circuits that each produces only a partial phase shift. For example, if we have N identical LC sections, as shown in Figure 12, we can design the section using: sin( Θ) Zos Ls = (1 + cos( Θ)) ω sin( Θ) Cs = Zos ω o o where: Θ = π 2 1 N RS Z0m Z0m Vps R=Z0m RS Z0m Z0m Vns R=Z0m Ls Ls Ls 2*Rpar Cs M Cs Ls Cs M Cs Ls Figure 12 Improving manufacturability by using multiple LC sections and mutual inductor coupling Cs Cs M Ls Even better use of the circuit comes when we take IO imperfections into account. Beside the capability to filter common noise, this is the circuit s best property. An IO usually has output capacitance that effectively band-limits the driver. This output capacitance is the major problem in high-speed IO design. It also reduces the output return-loss.

19 It is easy to see the effect of the parallel stub in this case. If the parallel stub is implemented within the chip on-die or on the package substrate, we can get rid of the transmission line between the driver IO and the parallel stub junction. The inductance of the parallel stub forms a parallel-resonant circuit with the output capacitance Co. This is illustrated in Figure 13. RS Z0m Co Ls1 Vps Cs1 R=Z0m M1 2*Rpar Z0m RS Co Ls1 Cs1 Vns Figure 13 Parallel stub can be used to resonate with output capacitance and improve the bandwidth of the driver As the parallel stub in Figure 13 is an integral part of the driver, the value of Rs can also be increased to improve the differential losses of the stub as well as to improve the differential return-loss. Several parallel stubs can be attached to the same nodes, as long as the frequencies that they need to filter are far apart, so there is no interaction. In the case of harmonics of the data rate, the frequencies are such that there will be no interaction between the parallel stubs. The only requirement is that all of the parallel stub load resistors must be treated as being in parallel, so they have to be adjusted. If an additional stub is attached, for example, we can double the load resistors to achieve the same performance. The parallel resonant circuit can help the differential return-loss as well as decrease the differential insertion loss, as it is obvious. It is easy to modify expressions derived earlier to plot the effect of the Co on the parallel stub. In Figure 14, we can see the differential S21 of the circuit with the parallel stub and compare it to plots from Figure 13. The first interesting and beneficial property is that the peak in differential S21 is moved from data rate to ½ data rate, which makes a better equalizer. It has a similar effect on the differential S22, as we achieve better return-loss at frequencies around the Nyquist frequency where there is more power in differential signal. A last inspection of Figure 17 gives simulated differential eyes of circuit with parallel stub (blue eye) and circuit without parallel stub (red eye) over the same channel. We can notice that red eye is not having any advantage in eye opening and at the same time it has more jitter.

20 One more reason why the parallel stub should be implemented as close as possible to the transmitter is that the parallel stub does not dissipate the common mode signal; it only reflects it back. The reflected common mode is then dissipated on transmit resistors. The closer the filter is to the transmitter, the less chances there are for the common mode to radiate into the air. Figure 14 Differential S21 of parallel stub with Co=1pF Figure 15 Differential Return loss of parallel stub with Co=1pF

21 Figure 16 Common mode insertion loss of parallel stub with Co=1pF Figure 17 Eye of circuit with parallel stub (BLUE EYE) vs. the circuit without the stub (RED EYE) over 10 inches of FR-4 at 10Gbit/s

22 Summary High-speed digital signals are usually carried over PCB boards, backplanes and cables using differential lines. Unfortunately, if these lines contain the common mode, this can present a source of radiation and can result in failing to meet EMI requirements. EMI problems usually arise due to common mode signals having strong components at frequencies associated with data rate and harmonics of data rate. If they reach effective radiators like connectors or inadequately shielded cables, these common mode signals are radiated and are responsible for failing EMI specifications. Common mode chokes, based on a high level of magnetic coupling, present a high impedance for the common mode and they are broadband, but they cannot achieve more than 10 db of attenuation in the frequency range of interest without seriously affecting the differential mode. When this is not enough, engineers, in some cases, try to use two common mode chokes in series to improve performance. Also, common mode chokes need to be mounted on the surface of the PCB and, due to that, require PCB vias in series that can seriously affect the differential signal. Series PCB vias are necessary in case of ferrite-based common mode chokes, as using transmission lines on the surface of PCB will also cause radiation. The common mode power spectral density is usually concentrated at discrete frequencies harmonically related to the symbol rate. It is beneficial to concentrate the attenuation of the common mode at these frequencies. Basically, it is preferable to attenuate the common mode only at symbol-rate harmonically related frequencies. By using narrow band-stop common mode filters, we can achieve more attenuation at the frequencies that most affect the EMI. Properties of common mode signals with respect to EMI are shown in the first section of this paper. This first section explains why narrowband common mode filters are preferable to common mode chokes. In second section, the new simple structure using a loaded parallel stub is presented. It can be used to filter the common mode noise that exists on transmission lines carrying differential high-speed digital signals. The theory of operation based on odd and even propagation modes is presented in this section. The structure can also be used as the analog equalizer. It can be easily used to equalize cables used in high-speed digital communication. Design formulas are presented that show how various levels of equalization can be achieved. The common mode narrowband filter center frequency and the amount of differential mode equalization are independent, which increases the versatility. Another application is to neutralize the excess capacitance along the differential transmission lines, like large via capacitive stubs on the backplane. References: 1. E.Kreyszig, Advanced Engineering Mathematic, John Wiley&Sons, W. R. Bennett, J. R. Davey, Data Transmission, McGraw-Hill, 1965.

23 3. K.C. Gupta, R.Garg, I.J.Bahl, Microstrip Lines and Slotlines, Artech House, Matthaei, Jones, Young, Microwave Filters, Impedance-Matching Networks and Coupling Structures, McGraw-Hill, A. Zverev, Handbook of Filter Synthesis, John Wiley&Sons, P. Acimovic, Novel Band-Stop Common Mode Filter for High-Speed Digital Data Transmission, DesignCon2007.