Mitigating Crosstalk

Welcome to the New Year!

My family and I have returned from a semester spent at the University of Oxford, where I tutored undergraduates, studied signal integrity issues on-chip, and delivered my High-Speed Digital Design seminar at several European venues. Travel is wonderful, but it is great to be home, too.

Those of you interested in the release of my long-awaited new book on signal integrity will be pleased to hear that the printing process is now complete.

  • Howard W. Johnson and Martin Graham (2003). High-Speed Signal Propagation: Advanced Black Magic. Prentice-Hall, ISBN 0-13-084408-X.

The new book a companion to the original book, (2003) High-Speed Digital Design: A Handbook of Black Magic. The two books may be used separately or together. They cover different material.

The following article is a continuation of my response from vol. 5, #11, "Acceptable Crosstalk".

In that previous newsletter, Mr. Manivassakam asked what level of crosstalk would be acceptable in his design. This note considers what can be done to reduce the amount of crosstalk in his board.

Mitigating Crosstalk

For crosstalk to perturb the operation of a state machine, the following sequence of events must take place.

  1. First, your system must generate an aggressive signal. This is usually an intentional signal, meaning that it appears in the system for some useful purpose other than just to bother the victim.
  2. The aggressive signal couples through some mechanism into the victim circuit.
  3. The coupled signal arrives at a time when the victim circuit is receptive to noise.
  4. The coupled signal exceeds the available noise margin in the victim circuit at the time of reception.
  5. In cases where the noise from one aggressor is not enough to cause data errors, the aggregate noise from multiple simultaneous aggressors may cause errors.

This five-point outline of the crosstalk scenario suggests five ways to reduce the impact of crosstalk.

(1) Shrink the Aggressor

This is not always possible with digital logic; however, you can adopt a policy of at least separating groups of nets according to their signal amplitude. This policy prevents large-voltage nets (e.g., 3.3-V) from affecting smaller-voltage nets (e.g., 1.5-V).

For example, in the design of a connectorized interface you might specify operating voltages for the synchronous data at a level lower than the asynchronous control signals.

(2) Reduce the Coupling

A solid reference plane is the most powerful tool available for reducing crosstalk within a pcb. The reference plane may be a ground plane or a power plane; either can be equally effective in reducing crosstalk between digital signals.

Once a solid reference plane is in place, the crosstalk between two parallel nets may be dramatically reduced by either spacing the nets further apart, or by pressing either trace closer to the plane.

If H represents the height of a trace above the nearest reference plane, and if the trace-to-trace separation S exceeds H, then crosstalk plummets roughly quadratically with increased separation. Doubling the spacing cuts the crosstalk to roughly 1/4 its original level.

Pressing the traces closer to the planes also reduces crosstalk, with the caveat that as the traces are pressed closer to the planes the trace width must be correspondingly reduced to maintain constant impedance.

Noise couples differently inside an integrated circuit. Increased separation does not necessarily decrease all forms of crosstalk (especially the substrate noise that occurs in a chip with a low-resistance substrate). Increased separation definitely reduces the direct parasitic capacitance from track-to-track, but since the metal tracks in a chip are held relatively far up away from any other bits of grounded metal the severity of the drop-off in crosstalk is not quadratic.

Differential signaling can reduce the coupling between nets, but the gains achieved are not as dramatic for pcb traces as for twisted-pair cables, because the pcb traces are not twisted.

Differential signaling works particularly well as a countermeasure against the changing ground voltages apparent on either side of a connector.

(3) Change the Timing

Regarding crosstalk onto a synchronous net, the crosstalk only matters within a small window around the moment of clocking. If you can adjust the clock feeding the aggressor so that it changes state at a time when the victim is NOT being clocked, then the interaction between the two circuits disappears.

Some mixed-signal chip architectures provide separate time slots for analog and digital processing, pinging back and forth continuously between analog and digital modes.

(4) Improve the Receiver Noise Margins

The noise margin available at the receiver is the difference, in volts, between intentional signal presented to the receiver and the internal threshold used within the receiver. An additive noise signal smaller than this margin cannot possibly create a data error.

Uncertainty in the exact setting of the receive threshold directly reduces the available noise margin, which in turn directly reduces the degree of additive crosstalk your system can tolerate.

For example, in a 3.3-volt system a driver going LOW might be expected to produce a signal no greater than 0.3 volts, while the receiver is guaranteed to respond to any signal less than 0.8 volts. The difference (the noise margin) in the low state is 0.5 volts.

Had the receiver threshold been guaranteed between 1.6 and 1.7 volts, the noise margin in the low state given the same transmitter would have been 1.5 volts, THREE TIMES BIGGER.

Differential receivers generally have very well placed thresholds, which partly accounts for their popularity in noisy systems.

In a highly attenuated system (like a 2.5-Gb/s backplane), any increase in the received signal level obtained by using a better transmitter (perhaps with pre-emphasis), or a less lossy transmission media, will directly increase the noise margin.

Most digital logic operates so far above the intrinsic internal noise floor of the semiconductor structures used to build the receiver that cooling the receiver doesn't help greatly, however in the design of some infra-red receivers the reduction in noise achieved by cooling the receiving elements is worth the cost of the cooling apparatus.

(5) Reduce the Number of Aggressors

While it is admittedly difficult to modify one's architecture in order to reduce the number of aggressors, it is in some cases possible to modify the number of aggressors that change state simultaneously.

This is accomplished in some systems by skewing the clock used on various banks of logic such that the aggressors change state at slightly different times. The crosstalk resulting from such a system displays a sequence of smaller blobs of crosstalk, rather than one grand superposition of all signals switching at once.

It may in some cases be possible to reduce the number of simultaneously switching signals with data coding. This approach is used in the coding of address lines to reduce the power required and noise generated when addressing memory sequentially.

I hope this structured examination of crosstalk has been helpful to you.

Best Regards,
Dr. Howard Johnson