Fixing Signal Integrity Violations

A signal integrity violation, whether it is a crosstalk delay that breaks setup or hold or a glitch that is too large, has one cause: too much coupling for the strength of the victim. The fixes are therefore of two kinds. You can reduce the coupling, or you can make the victim more robust. This article lists the usual fixes, shows how much a change of the coupling capacitance is worth on the demo circuit, and explains how to check that a fix has worked.

8 min read | Intermediate

What You'll Learn

• List the main ways to remove a crosstalk violation
• Explain which fix helps the delay and which helps the noise
• Estimate the benefit of a smaller coupling capacitance on the slack
• Check that a fix does not create a new violation

Reduce the Coupling

The most direct way is to make the coupling capacitor smaller. The capacitance between two wires falls when the distance between them grows, and rises with the length over which they run in parallel.

• Spacing: route the victim and the aggressor farther apart. The coupling capacitance falls quickly with the distance, and the tools can add spacing on the nets that need it.
• Shielding: put a wire at a fixed voltage, a supply or a ground, between the two nets. The coupling to the aggressor becomes a coupling to the shield, which does not switch. The capacitance does not disappear, but the Miller effect does: the victim sees a quiet neighbour.
• Shorter parallel runs: change the routing so that the wires cross instead of running side by side, or use different layers for the two nets, with a layer in between that has other signals.
• Different tracks: reorder the wires so that a sensitive net has quiet neighbours on both sides.

Strengthen the Victim

A strong driver pulls the victim back faster and keeps it near its value while the aggressor switches. For delay, a stronger driver makes the victim edge less sensitive to the added capacitance. For noise, it reduces the height and the width of the bump.

• Upsize the driver of the victim. It is a quick fix, but a larger cell has a larger input capacitance and uses more power.
• Insert a buffer in the middle of a long victim wire. Each half is shorter, has less resistance and less coupling, and the driver of the second half is a new, strong cell.
• Shorten the victim wire by moving the driver and the load closer together.

Soften the Aggressor

• Downsize the aggressor driver, or add a small resistance, so that its edge is slower. A slower edge couples less.
• Change the logic so that the aggressor does not switch at the time of the victim. This is a design decision, supported by the timing windows of the article about aggressor alignment.

What a Smaller Coupling Is Worth

The demo circuit shows the size of the effect. The slow-down case, with 0.006 pF of coupling seen by the victim, gave a setup slack of 1.0624 ns on the path between the two flip-flops. These are the values for a smaller capacitance, read with the quiet-neighbour assumption:

Each step gives some setup slack back: 0.0889 ns for a shield or for non-overlapping windows, a further 0.0396 ns for double spacing. Notice the hold column: it goes the other way. The hold slack falls as the coupling is removed, from 0.6998 ns to 0.5946 ns, because less capacitance makes the data faster. A fix for a setup problem reduces the hold margin of the same path. After every change, both checks have to be run again.

Checking the Result

• Run the full analysis with crosstalk after the fix. The change of the capacitance changes the windows of other nets too.
• Look at the neighbours. A wire that moved to make room for the victim is now closer to another net.
• Check both setup and hold of the paths that went through the net.
• Check noise as well as delay. A stronger driver reduces both, and spacing reduces both, but a buffer changes the timing of the whole path.

Order of Work

In a typical flow the tool fixes violations automatically in the routing and optimization steps, with spacing, shielding and sizing. The designer looks at the ones that remain, which are usually on a critical clock, in a region with high congestion, or on a net with a special requirement. For those, a manual change of the routing or of the floorplan is the right answer. Clock nets are often routed with extra spacing or shielding from the start, because a glitch or a delay shift there affects many flip-flops.

Common Mistakes