Double Tracking

Some of you know that I live way out in the country in the rural eastern portion of Washington State, high in the Cascade mountains, in a dry, sunny region called the Methow Valley. I occasionally get questions about what it is like to live so far from the big city, away from the hustle and bustle of daily city life. In response to those questions, I have attached at the end of this letter a brief story about what happened one day here at our ranch in the town of Twisp, Washington.

Double Tracking

Let's begin this discussion looking at the belt-and-suspenders, super-safe differential stripline architecture. The application I have in mind is a backplane, using ZD-style backplane connectors. The backplane runs horizontally (as in a standard rack-mounted application), and has about ten connectors.

Figure 1 shows a partial layout for one connector in plan view. This snippet of the layout shows 9 wafers, each holding four differential pairs. At the bottom, I show the G  S+S–G  pin configuration, indicating that the differential pairs (S+S–) are separated by ground pins (G). The traces draw proceed to the right, headed toward the next connector.

Figure 1—This 9-wafer ZD-style backplane connector layout accommodates 36 differential pairs, four pairs per wafer (figure courtesy of ERNI GmbH).

Some engineers would call this a four-row connector, because it holds four pairs in each column. Others might call it an eight-row connector, because it has eight signal connections, or perhaps a 13-row connector, because that is the number of pins in each column (including grounds). When you order connectors, make sure you and your vendor are using the same terminology. Also, note that in this drawing each column, or wafer, extends horizontally while rows extend vertically (strange, but that's the convention). In the clearest terms, this layout supports four pairs per wafer and shows 9 wafers.

I would like to produce an escape pattern that routes all four differential pairs away from the connector using only two signal layers. I call this a double-track differential escape pattern, because it provides escape paths for two differential pairs (four wires total) between each wafer. If I can accomplish the double-track escape, I can escape the whole connector (in one direction) using only two signal layers. The entire backplane will of course require many more than just two signal layers, because it must allow some signals to pass through my connector without touching it, and also account for the crossing of signals at intermediate locations.

One possible double-track escape pattern appears in Figure 2. The figure shows a cross-section view of the layer stack. The connector's signal-pin vias appear on either side (only partially visible, in yellow). The solid planes above and below represent the minimum width of the web of continuous reference plane material existing between the columns of signal pin vias.

Figure 2—(Cross-section view of layer stack) This double-track (two pair) routing scheme passes two differential pairs between each column (wafer) of connector pins.

This pattern uses a 6-6-6 differential pair configuration. That notation indicates an edge-coupled differential pair comprising a 6-mil trace, with a 6-mil space, followed by another 6-mil trace. In this drawing, possible locations for the connector-pin signal pads are indicated in light blue. Some of the signal pads may be missing, however, we must still provide a 6-mil minimum clearance for our traces in case those pads are present.

The available space between the blue pads is 57 mils wide. I have filled that space with two differential pairs (red), observing the 6-mil requirement on both sides, and leaving a 9-mil gap between pairs. The vertical space required to produce a 100-ohm differential impedance with this configuration equals 16 mils above and below the traces, for a total thickness of 32 mils.

The near-end-crosstalk coefficient (NEXT) for this configuration is 0.031, or about three percent crosstalk pair-to-pair. Keep in mind that the total delay of the coupled portion of these pairs will be on the order of 127 ps (in FR-4). Any edges faster than 127 ps will generate fully developed NEXT waveforms of three percent. You will probably also encounter additional crosstalk from various other connector pads along the way, since they approach so close to the traces. If amount of crosstalk from this layout seems OK for your design, this is not a bad way to go. The layout is simple and uses few layers.

The systems I have worked on recently cannot tolerate three percent crosstalk. If you are working on a serial backplane, you probably cannot tolerate that much, either. Consider that one pair, threading through ten connectors in a row, picks up three percent crosstalk underneath each connector it traverses. Those crosstalk numbers can add up quickly.

So, let's try a different layout. Figure 3 gives up on the double-track idea. Each signal routing layer places only one signal pair between the wafers. This is a bulletproof design, with great appeal. Crosstalk from pair to pair will be (as far as we digital folks are concerned) non-existent. An analog guru might be able to measure the crosstalk leaking through the connector via holes in the reference plane, but not me. This design works. Notice that we did not have to quite double the board thickness to do it. The double-track design used very tightly coupled pairs, thus requiring either incredibly tiny, thin traces or a generous spacing between the planes. This single-track design accommodates a more widely spaced pair. This 6-8-6 layout is less heavily coupled, and thus for the same trace width can make do with more tightly spaced planes. The plane spacing here is 10 mils above and below. Both designs produce a differential impedance of about 100 ohms.

Figure 3—This single-track (one pair) routing scheme passes only one pair between each column of connector pins. Crosstalk is practically non-existent.

If you saw my previous newsletter vol. 7 #04, "Squeeze Your Layer Stack", you know that I like thin backplanes. Keeping the 6-8-6  trace configuration fixed, you can slim down the layer stack in Figure 3 by deleting every other solid ground layer. That deletion raises the impedance of the differential pairs. Jamming the solid planes closer together then brings the impedance back down to 100 ohms.

This change still leaves at least one good, solid return path underneath every differential pair, and it really slims down the board, but there is a hitch: the crosstalk re-appears. The design in Figure 4 shows the numbers. With the center ground plane missing, you get a total thickness of 30 mils (as opposed to 40 from last time), but the crosstalk pops right back up to +3.3 percent, assuming the differential orientation as drawn. Darn. It didn't help (yet).

Figure 4—This routing scheme uses fewer layers, and is more vertically compact, than the single-track approach. Crosstalk (NEXT) is 3.3% pair-to-pair.

To fix the crosstalk problem you might try offsetting the pairs from each other. Figure 5 shows the result. The horizontal offset is 25 mils, the most we can have while still meeting our 6-mil clearance requirement on each side. The NEXT crosstalk, pair-to-pair, falls to only –1.1 percent, a better value, but perhaps still not good enough. I want it even lower. What can I do?

Figure 5—Setting the offset to x=25 mils, the maximum available in this layout, reduces the crosstalk (NEXT) to a level of –1.1 percent pair-to-pair.

Reviewing the numbers from Figures diagrams 4 and 5 closely, I see that the crosstalk in Figure 4 is positive, while that in Figure 5 is negative. This happens because in Figure 4 the traces line up + to + and – to –, so you of course get a positive NEXT coefficient. In Figure 5 we have offset the traces so severely that now the + and – traces line up most closely, generating a negative NEXT coefficient. As we slide the traces from the Figure-4 position to the Figure-5 position, nature behaving in a continuous manner, shouldn't we at some point obtain a NEXT coefficient of zero?

Surprisingly, that is exactly what happens. There exists a natural null in the pair-to-pair crosstalk for pairs on adjacent layers, but it only happens if you choose precisely the right offset. Figure 6 shows the position of that null. This solution was obtained empirically, using HyperLynx LineSim V7. The plot of NEXT crosstalk versus offset in Figure 7 shows that the null is fairly well behaved (i.e., it is not a narrow, deep notch). A layer-to-layer offset tolerance of 0.001 in. should allow you to reliably contain the pair-to-pair crosstalk to a value below 0.003. In my opinion, that's good enough. The net result is a board 25thinner than Figure 3, using only four layers instead of five.

Figure 6—The crosstalk (NEXT) varies as a function of the pair-to-pair offset, x. The particular offset drawn in this geometry produces zero crosstalk.

This null reminds me a similar effect used in some German LAN cables. These cables (popular in the 10BASE-T era) contained four wires, held in a rigid "X" profile (one wire at each tip of the X). Labeling the wires clockwise as RED, YELLOW, GREEN, and BLACK, you wired opposing wires into pairs (RED and GREEN, or YELLOW and BLACK), to make up a 10BASE-T Ethernet cable. Because the wires were held in a rigid exact (and highly symmetrical) position they existed in a kind of crosstalk null much like the one produced in Figure 6.

Figure 7—Near-end crosstalk (NEXT) between two differential pairs implemented on adjacent layers passes through zero as it goes from positive (on the left side of the figure, representing traces piled right on top of each other) to negative (on the right side, representing widely spaced pairs).  

Best Regards,
Dr. Howard Johnson

Questions & Comments

All students who attend High-Speed Digital Engineering Week seminars at Oxford University have the opportunity to talk directly with a variety of experts about all aspects of signal integrity and high-speed system performance. We'll have independent experts there available to answer your questions about system design, high-frequency problems, EMC, measurements, noise, system design and simulation workflow.

The very first High-Speed Digital Engineering Week event held at the University produced comments like these:

  • A very good exchange of ideas. I've learned a lot...,
  • [The session] …Covered a good range of issues. Good to hear different viewpoints..,
  • [The session was] …Informative and useful. It was good to hear other peoples' "real" issues and their solutions..,
  • Very useful to see a variety of opinions on how to solve various problems.

Ninety seven percent of the attendees said they would recommend it to their colleagues.


Personal log, dog-date: "First Day of School"

My dog Val chose the first day of the new school season to teach herself an important lesson. Val is a healthy, vigorous, four-year old Australian Shepard. A swirling mass of brown, black and white fur, this breed of dog relishes the outdoor life; constantly zipping between the barn, house, and the water hole. If something moves within the visible horizon, she and her twin sister Blackie are right on top of the action.

When I'm packing my truck for a day's outing in the pastures, Val always appears, like a genie, smiling and panting, in the bed of the pickup. How she vaults over the side of the vehicle I will never know, as it happens too quickly to observe with human eyes.

On this particular morning, Val and her litter-mate woke early and threaded their way downstairs past piles of new-school materials to the front door for their morning constitutional. The stacks of notebooks and paper and backpacks had been all neatly organized and prepared by our daughters, Katy and Allie, the night before so as not to be late for the first day of school. Allie, 12 years old, going into Junior High for the first time, insisted especially that there be no last-minute delays leaving the house on this crucial morning.

Out on patrol, the dogs make their way, sniffing and barking, checking tracks left by the bear, deer, coyote, and what-have-you that tramped through the yard during the night-time hours. Inside, the girls wolf down their cereal and brush their teeth. Allie straightens her new, bright-white, first-day-of-school blouse.

Suddenly, a round of furious barking shatters the early-morning calm. This is not an unusual occurrence at our house. The dogs could have turned up anything—a mouse, a skunk, a man on horseback—regardless the source they dutifully obey an ancient instinct the let everyone know they saw it first.

Within moments, the barking turns into yelps and snarls; an epic battle rages outside.

The children run out to see and come back screaming and in tears. Then I hear my wife, Liz, add her shouts to the ruckus. I, oblivious to the commotion, being quite accustomed to the increasingly frequent emotional outbursts of my pre-teens, merely focus my attention more forcefully on my first morning's project: the surgical extraction of the last remaining portion of a small splinter jammed in my thumb.

When the level of general panic swirling about the house reaches a level that even I cannot ignore, I trundle downstairs to the front porch, curious to see what is causing all the excitement. There, spread before me, lies a tableau of pain and suffering. I hear the screams of girls and the snarls of dogs. A high level of general hysteria prevails. Liz sits with pliers in hand. Blood is splattered in every direction on new school clothing. In the center of the hubbub, defiant, even arrogant, apparently impervious to pain, still barking, stands Val, the wonder dog, looking particularly fearsome as she now sports a hideous array of spikes emanating from her snout.

My first reaction is one of disbelief, and early-morning confusion. I don't remember Val having spikes growing out of her head, although they do look like they would deter the advance of any rational-minded creature wanting to avoid getting stuck. Slowly it dawns on me that the spikes point the wrong way, blunt ends sticking forward. Val has, apparently, encountered her first porcupine.

It takes a lot of patience to pull porcupine quills out of a dog's snout. First, you have to catch the dog. Then you must calm her down and then slowly, gently as you can, you grasp each quill one at a time with pliers and YANK. Dogs like this even less than kids like doing homework.

After all the excitement died down, the kids packed their stuff in Liz' car (more than a little late) and left for school. As they all sped away down the hill, with admonitions from the children to never let the dogs go outside ever again, I found myself hoping, with the eternal optimism of parenthood, that somehow, during the coming school year, the children would learn a few lessons even half as valuable as the lesson Val learned on that first day of school.

H. Johnson
Twisp, WA
Sept. 08, 2004