首页资源分类PCB layout > DesignCon2015:A Cure for Intra-pair Skew in High Speed Differential Signals

DesignCon2015:A Cure for Intra-pair Skew in High Speed Differential Signals

已有 445500个资源




    标    签:skew

    分    享:



    1. Skew 对差分信号的直接影响并不大, 关键是会容易转化成共模噪声, 而且这个转换时可逆的, 也就是skew 会增加差模转共模的能力,也因为更容易受共模噪声干扰。

    2. Skew一般很那定义和测量,作者提出一个方法:用共模噪声来衡量skew, 0 共模噪声代表0 skew,可以简化问题,方便评估和测试。

    3. 提出一个共模滤波器,可以抑制共模噪声,但作者自己也说这个方案还不完善。



    DesignCon 2015 A Cure for Intra-pair Skew in High Speed Differential Signals Mike Jenkins, Xilinx Abstract As serial data rates increase, the effects of small PCB layout imbalances and fiberglass weave inhomogeneity cause skew between the positive and negative signals in the pair which is an increasing fraction of the symbol width. This contributes to both increased jitter and increased common mode energy. Small cutouts in the current return ground planes for these signals can create common mode band-stop filters that significantly reduce these deleterious effects while passing the differential signal virtually unaffected. Author Biography Mike Jenkins received his BS in Electrical Engineering and MA in Mathematics from the University of Illinois (Urbana) and his MS in Electrical Engineering from Syracuse University. During his 40 year career, he has held engineering positions at IBM (Data Systems and General Products Divisions), LSI Corp., and Xilinx Corp. with a primary focus in the areas of signal integrity and SerDes design and analysis. He holds 18 patents. Introduction Signal rate increases are outpacing improvements in the materials comprising the transmission channel. Decreases in channel length have not kept pace, either. Consequently, mechanisms that cause skew between the positive and negative paths of the differential signal (intra-pair skew) are becoming a more urgent problem. Viewing this situation as undesirable differential-to-commonmode conversion rather than as differences in delay (i.e., skew) will be shown to attack the problem more directly and to yield a simple solution – although one that violates one of the most sacred dictums of PCB signal integrity: do not run high speed lines over split ground planes. The Effects of Intra-pair Skew While the most obvious concern is whether the received differential signal may be directly degraded by intra-pair skew, differential signals are quite robust against moderate amounts of skew (i.e., up to about +/- 0.5 UI). Within this range, the most worrisome effect is mode conversion, that is, creation of common mode voltage from what had been a purely differential signal. Common mode voltage will result in radiation causing electromagnetic interference (EMI) much more strongly than purely differential signals from which the radiated fields largely cancel. This also results in three follow-on negative effects:  First, the presence of a common mode component will disadvantage any receiver with a less-than-perfect high frequency common mode rejection ratio (CMRR).  Second, in any passive channel, any energy converted to common mode reduces the energy in the differential signal. That is, the eye opening at the receiver will be reduced.  Third, for any passive, linear, time-invariant channel where intra-pair skew occurs, reciprocity applies. This means that any channel which converts differential signals into common mode is equally capable of converting common mode into differential signals. A channel that can radiate common mode from a nominally differential signal is just as capable of receiving it, consequently suffering differential mode interference. The Causes of Intra-pair Skew Perfect symmetry between the positive and negative signal in a differential pair is degraded both by physical layout asymmetries (sometimes unavoidable) and by random variations in PCB fabrication. An example of physical asymmetry is shown in the top left of Figure 1 where the location of pins in a connector footprint induces an unequal layout of the positive and negative traces. Bumps, of course, can be added to lengthen the shorter trace, but this is only an approximation of symmetry. The drawing in the top right of the same figure depicts various random copper etching imperfections that can occur in any trace, degrading symmetry. A particularly worrisome random imperfection is described in the bottom of the figure. The fiber glass used in PCB construction has a different dielectric constant than the resin that fills around it. Therefore, a trace that happens to run over a fiber will propagate at a different speed than one that runs primarily over the resin. Since the fiber pattern cannot be registered relative to the etched copper, this is a random process. The figure depicts the nightmare scenario where one trace is completely over the fiber and the other lies between fibers. Since the difference in relative dielectric constants can be 0.5 or more [1, 2], four inches of PCB trace with the worst possible alignment to the weave could induce intra-pair skew greater than one bit of a 28 Gb/s signal. To counter this, fiberglass weaves are increasingly designed to homogenize the dielectric constant. Figure 1 Various effects causing intra-pair skew Zero Common Mode Implies Zero Skew In the general case, skew can be a difficult quantity to define and to measure. If the two waveforms are not of identical shape, then the value of skew will change depending on where the threshold is set. Attempts to pin down this measurement procedure unambiguously can involve some questionable assumptions. Also, attempting to solve this problem by adjustable delays can result in relatively complex, power-consuming circuitry [3]. Further, delaying one side of a pair to equalize their threshold crossing time does not, in general, guarantee zero common mode. For the limited case of intra-pair skew in differential signals, instead viewing this problem as undesirable common mode (which is the major problem resulting from intra-pair skew) leads to much simpler definition of a figure of merit as well as much simpler solutions. If one can remove all common mode from a differential signal (Vpos - Vneg), then Vpos + Vneg = 0. So when Vpos=0, Vneg=0, as well. That is, there is no skew. A Simple Common Mode Filter The geometry in Figure 2 below evolved heuristically from an experiment years ago using some semi-rigid coax soldered to a single-sided copper-clad board with a slot etched across it. Later, some short wires were soldered across the slot to observe the effect on the common mode transfer function. The “H” shape represents a void in the GND plane both above and below the strip line differential pair. The most significant dimensions, L1 and L2, are related to the center frequency of the common mode stop band by: L1, L 2 = /4 = [300 mm/ns] / [4 * sqrt(r) * fNyquist] Where 300 mm/ns is the free space velocity of propagation, r is the relative dielectric constant, and fNyquist is the Nyquist frequency (i.e., half the data rate – at least for binary (PAM2) signaling). This equates to roughly 150 mils for a 20 Gb/s signal rate and a typical PCB dielectric. The values for W and Gap are roughly 2 to 3 times the dielectric thickness to avoid significant impairment of the signal propagation (W) or significant coupling across the Gap. Figure 2 A simple common mode filter The basic principles is that the “diving board” shapes in the GND plane (L1 and L2 in length) comprise two cascaded filters, resonating at fNyquist for any common mode signal components, causing a coupling between the two signal conductors (dashed arrows in Figure 3) which would be absent without the GND plane cutouts. Staggering L1 and L2 fractionally above and below the nominal target is an attempt to widen the aggregate stop band of these two cascaded filters. More complex cutout geometries have been described [4]. The effect of this construction when the two inputs are skewed is best seen in the time domain. In Figure 3, the waveforms are derived from inverse Fourier transforms of measured s-parameter data. The top pulse response is the “normal” one, representing the insertion loss along each signal path in the pair (typically loosely coupled). The bottom pulse response is the forward coupling between the traces in the pair1. For differential signaling the input pulse on node 3 is negative, so the output pulse at node 2 is the top pulse response minus the bottom (coupled) pulse response. Hence, the coupled pulse increases the composite output pulse prior to the 1 In the absence of the ground cutouts, TEM wave propagation in such a homogeneous medium produces no forward coupling, since the inductive and capacitive components exactly cancel. dashed red line in the figure (increasing the pulse amplitude) and lowers it to the right of the dashed red line (sharpening the falling edge). Figure 3 Cross-coupling induced by common mode filter The magic occurs when the inputs at nodes 1 and 3 are skewed. Say, for example, that the input pulse at node 3 arrives later than the input pulse at node 1. In that case, the output pulse at node 2 is comprised of the top pulse minus the bottom (coupled) pulse shifted later in time. This causes the peak of the composite pulse at node 2 to occur later in time, minimizing the apparent skew relative to the output pulse on node 4 (as well as causing some undershoot on the falling edge). Analogously, the output pulse at node 4 (inverted when part of a differential signal) is comprised of the top pulse minus the bottom (coupled) pulse shifted earlier in time. This causes the peak of the composite pulse at node 4 to occur earlier in time, again minimizing the apparent skew relative to the output pulse on node 2 (as well as causing a bit of a bulge in the tail of the pulse). As the above paragraphs admirably motivate the old adage that a picture is worth 10,000 words, please consider Figure 4 depicting the result with and without the common mode filter on a 16.7 Gb/s differential pulse transmitted through 6 inches of Megtron 6 strip line. One input signal in the pair is delayed relative to the other by +½, 0 and -½ bit time top to bottom in the 3 plots. In each, the dashed green curve is the composite output pulse at node 2, and the dashed blue curve is the composite output pulse at node 4 (inverted for easier comparison). The solid green curve is the differential pulse output, and the solid red curve is the common mode output. Please note that in each case, the differential pulses with or without the filter are virtually identical. Also, please note that, in each case, the common mode output is considerably smaller with the common mode filter and that, in all cases, the two peaks of the common mode output without the filter are separated by one bit time which would generate a strong common mode tone at the Nyquist rate. Figure 4 Differential and common mode pulse responses vs. input intra-pair skew. Dashed curves are the individual positive and negative (inverted or comparison) output pulses. Solid green curve is the differential output pulse, and solid red curve is the common mode output pulse resulting from the input skew. The differential output pulse responses are examined in more detail in Figure 5. The top graph compares the pulse shapes with the common mode filter (solid curves) and without it (dashed curves) for skews of -½, 0 and +½ bit time. The pulse shapes exhibit some widening and lowering of peak amplitude at the extreme skews. However, these effects occur equally with or without the common mode filter. The bottom graph in Figure 5 magnifies the tails of these pulses. Here it can be seen that the common mode filter disturbs the smoothness of the tail relative to the case without the common mode filter. Typically, a continuous time linear filter attempts to equalize these tails, so any “bumps” will remain as a residue. However, these variations are small. (Each vertical division in this graph is 0.25% of the input differential pulse amplitude.) In fact, a small reflection occuring on the hardware without the common mode filter at about 113 UI (dashed curves) is significantly larger than the ripple caused by the common mode filter. Figure 5 Pulse response shapes (top) and details of pulse tails (bottom) vs. skew with and without common mode filter Variations on a Theme Several variations on the cutout geometry have been tried. The results are summarized in Figure 6 below where the figure of merit is the bandwidth and magnitude of the common mode suppression, defined as the ratio of the common mode insertion loss to the differential mode insertion loss.  The base design, used in the results reported above, is shown at top right, exhibiting about 5 dB of suppression across a bandwidth roughly half the center frequency.  A parabolic shape (top, center) is an attempt to widen the stop bandwidth and shows promise.  Including the GND cutout on only one of the two GND planes in the strip line cross section (bottom, left) has very little effect.  Cascading two sets of GND cutouts (bottom, center) results in what one expects from cascaded filters. (Note the doubled common mode suppression.)  Unequal cutout lengths near the positive and negative traces (bottom, right) attempts unsuccessfully to widen the stop band. This also apparently induces an asymmetry between the positive and negative traces, causing common mode conversion. (See next section.) Figure 6 Common mode suppression for a variety of ground plane cutout geometries Isn’t This a Slot Antenna? Well…yes…probably. In fact, initial literature searches on “defected ground planes” yielded primarily papers on the subject of antennas. While the author does not claim any expertise in antenna theory, a slot antenna is described in the literature as a long thin opening in a large grounded conductive structure. When this structure is fed by a voltage across the slot (usually at the center), circumferential currents are induced around the slot, resulting in transmitted EM waves with a center frequency wavelength of twice the slot length [5]. If the three layers (GND/signal/GND) that comprise the common mode filter are buried in a PCB stack up, then most, if not all, of this radiated energy will be captured. The experimental results in Figure 7 below show the common mode insertion loss of a trace pair with a CM filter (solid line with notch) and without a CM filter (smooth, sloping solid line). The dotted lines show the sums of transmitted and reflected common mode energy for both pairs which are equal within measurement error. This confirms that the common mode energy that is not transmitted is actually reflected back to the source where it is absorbed rather than being lost to EM radiation. Figure 7 Transmitted plus reflected common mode power with and without CM filter One more related concern remains: If GND planes are slotted above and below a differential pair in order to block common mode, what about signals on the other side of those GND planes? The experiment shown in Figure 8 was laid out to create just that geometry, and measurable coupling from the differential pair between the GND planes to the pair below the GND plane was observed. The plots below show the differential-to-differential coupling (near-end and farend). The top/bottom asymmetrical layout of the cutouts in this experiment was an attempt to create a wider common mode stop band. In hindsight, however, this represents an asymmetry on the signal path that will produce mode conversion, generating common mode from an otherwise perfect differential signal. Further, the lobe reaching -30dB in the left plot (NEXT) is at 12.5 GHz, exactly the center frequency to which the left half of the ground cutout is tuned. In the right plot (FEXT), the prominent lobe at 10 GHz corresponds exactly to the center frequency to which the right half of the ground cutout is tuned. Figure 8 NEXT and FEXT coupled to differential pair on the opposite side of a ground plane cutout Summary Intra-pair skew is widely acknowledged to be an increasingly pressing problem at higher serial data rates. Skew is caused both by nominal component and layout geometries as well as random manufacturing variations, particularly the “fiber weave effect”. While skew does directly impact the received differential signal, its impact is only modest up to about +/- ½ bit period. The more immediate impact of skew is an increase in common mode voltage, increasing the transmission of EMI as well as susceptibility to EMI. Approaching the problem as one of excess common mode both simplifies the problem and directly attacks the main issue. Further, removing all common mode components guarantees zero skew between the positive and negative signals comprising a differential pair. A method of creating common mode band stop filters has been described using cut-outs in the ground planes of a PCB strip line. Measurements have confirmed their efficacy. However, this technique has been developed heuristically, so a number of questions remain, and more work needs to be done to predict not only the depth and width of the band stop, but to understand any interaction with other circuit elements. References [1] Bogatin, E., “Glass Weave Skew Problem May Be Solved,” DesignCon Community, 2013 [2] Simonovich, L., “Practical Fiber Weave Effect Modeling,” White Paper-Issue 3, Lamsim Enterprises, 2012 [3] R. Fung, R. Senthinathan, N. Chan (Ati Technologies Ulc). Intra-pair differential skew compensation method and apparatus for high-speed cable data transmission systems. US Patent 8286022 B2, January 12, 2009. [4] Yangyang Pang, Zhenghe Feng, “A compact common-mode filter for GHz differential signals using defected ground structure and shorted microstrip stubs,” 2012 International Conference on Microwave and Millimeter Wave Technology (ICMMT), Volume: 4, Publication Year: 2012 , Page(s): 1 – 4 [5] Slot Antennas (2010). Retrieved October 27, 2014, from http://www.antennatheory.com/antennas/aperture/slot.php

    EEWORLD下载中心所有资源均来自网友分享,如有侵权,请发送举报邮件到客服邮箱bbs_service@eeworld.com.cn 或通过站内短信息或QQ:273568022联系管理员 高进,我们会尽快处理。