Crosstalk is bad. How much crosstalk is too much? The only answer is... "it depends." In digital applications, most specs recommend less than -20dB for a signal-to-noise ratio. For example, crosstalk equivalent to 5 percent is -26dB.
In tightly spaced 50Ω edge-coupled stripline busses, the near-end crosstalk can be as high as 6 percent between adjacent lines. This could result in 12 percent near end noise in a bus. The solution is to space the lines farther apart. This is the origin of the rule of thumb, to keep the spacing between adjacent traces at least 2x the line width so that line-to-line near-end coupling is less than 2 percent and total near-end crosstalk is less than 5 percent.
With two routing layers between return planes in a dual or offset stripline stack up, there is another source of coupling: inter-layer coupling. If two lines on adjacent layers are routed parallel to each other, and on top of each other -- broad side coupled -- the near-end crosstalk can be much larger than when edge coupled. The figure below shows the calculated saturated near-end noise between two 5 mil wide, 50Ω transmission lines as their overlap shifts.
Broadside coupled near-end crosstalk as the overlap between two co-parallel traces is shifted.
When they are fully broadside coupled, the saturated near-end noise can be as much as 18 percent in some cases. This is huge.
Of course, this is why itís a good habit to always route lines on adjacent layers orthogonal to each other. The overlap is very small, and the signal-return current loops do not overlap. Inter-layer cross talk is dramatically reduced.
When routing density is at a premium, itís irresistible to use all available routing channels to get into the center of the BGA escape region, for example, by routing all available layers in the same direction. When this happens, lines on adjacent layers may be broadside coupled.
But what if we just keep the overlap region really short? If the overlap region is short, the total crosstalk will be less. While true, this is a case where putting in the numbers yields a very surprising result. How much overlap distance can we tolerate to keep the inadvertent broadside coupled near-end crosstalk to less than 5 percent?
The near-end noise will saturate when the time delay of the coupled region is 0.5 x the rise time. If the coupled length can be made shorter than this, the near end noise will decrease linearly with coupled length.
If the saturated near end noise is 15 percent, we would need the time delay of the coupled region to be less than about (1/3 x 1/2) = 0.17 x the rise time to keep the broadside coupled near end noise to be less than 5 percent.
If the rise time of the signal were 0.3nsec, as with a DDR3 signal, the maximum time delay of the broadside coupled region should be kept less than 0.3nsec x 0.17 = 50psec. With a typical signal speed of about 6 inches/nsec, this is a max coupled length of about 300 mils.
This means in the BGA escape regions, you better keep the inadvertent broadside coupled overlap shorter than 300 mils; otherwise, there could be more than 5 percent near end cross talk between adjacent signal lines. This is manageable.
But what about when the rise time shrinks, as with 2.5Gbit/s PCIe gen I signals? If the rise time is 0.1nsec, the maximum length to keep inadvertent broadside coupling drops to 100 mils, and if the rise time is 0.05nsec, as in PCIe gen II, now the maximum overlap region needs to be shorter than 50 mils. Here lies the challenge.
I know everyone in the DesignCon community is aware of this sneaky crosstalk path and knows to avoid it; however, I see a lot of other designs that worked just fine at 500Mbit/s but when scaled to 5Gbit/s, inadvertent broadside coupling causes the board to fail.