Predicting Skew for High Frequency, High Data Rate Designs

Controlling glass weave-induced skew has been a longtime concern for the PCB design industry. Initially, there was no method for doing so and, then as the issue and the problems associated with it evolved, some advancements were made. But, as has been noted on our website as well as several other places (such as the SI reflector), today’s high frequency, high data rate designs have brought skew control front and center among PCB criteria. This article will address the history of mechanically spreading glass weaves and how it successfully eliminated ragged holes in blind vias; then it segues to skew; how it impacts PCB performance; the steps taken within the industry to address it thus far and how, for the first time, it’s possible to predict and control skew at the forefront of the PCB design process.

Defining the Problem and A Little History

For the purposes of this article as well as today’s high-frequency, high data rate designs, the issues with skew are relative to differential signal pairs. Skew occurs when a trace runs over a glass fiber for some distance and then between glass fibers. When this happens, the dielectric constant varies substantially, causing significant variations in impedance and velocity along the trace. When the two members of a differential pair travel over the weave pattern these variations may not be equal in both members. This results in skew between the two sides of the pair and degradation of the signal at the receiver. And, as noted previously, the tight design constraints that we use in today’s PCB products mean that there is little wiggle room for error. Thus, unpredicted, uncontrolled skew which may even be minimal in form, can quickly become one of the hidden “gotchas” and can render a final product unmanufacturable, inoperable and unreliable. This is turn, creates real havoc with product market windows, competitive advantages and long-term profitability of the end product.

Before we started having issues with weave, the choice of the laminate and the glass style incorporated into it was based purely on cost. And, the laminate/glass weave style combination that was ultimately selected was always the one that cost the least. As a result, depending on the final cores, there came to be some default standards. They include:

• For two mil cores, the glass weave of choice was 106.
• For three mil cores it was 1080.
• For four mil cores, 2116 or 3313.

 The foregoing sufficed until the use of laser-drilled vias came on the scene.

 With the advent of increasingly smaller mobile phone devices and ultra-thin laptops, smaller geometry devices with pin spacings of less than 1mm became commonplace and they, in turn, led to the predominance of blind vias. And, that’s when we ended up with some very ragged holes such as those shown in Figure 1 when the 1080 glass weave was used.

The ragged hole shown in Figure 1 occurred as a result of the irregular distribution of the glass and resin within the weave of the 1080 core. In order to burn through the glass weave, the laser burned away too much resin resulting in the ragged hole.

Further examples of glass weaves are shown in Figure 2 which pictures 106 and 1080 glass cloth. Note: the transmission wire running across the weaves is described later in this article relative to differential pairs.

Here the large voids, which are comprised of pure resin, are readily visible. When the power of the laser beam is set high enough to burn through the glass, in the regions where the bundles are pure resin, the beam drills through the copper pad undercutting the resin on the hole sides. The result is a ragged, unreliable hole.

The solution to the foregoing problem was to mechanically spread the glass so that it is even and flat thereby reducing the chance that ragged holes will be created. And, some years back, within the IPC—the Association Connecting Electronics Industries—there was a committee to specify the mechanical spreading of glass. And, with the focus on blind vias, significant progress was made. Figure 3 depicts the mechanically spread replacements for glass weaves shown in Figure 2. 1067 glass is the mechanically spread replacement for 106 glass while 1086 glass is the spread replacement for 1080 glass. As can be seen, the weave is flat and there are no visible resin voids.

What About Differential Signals?

Per the above, there’s been significant progress made relative to spread glass and the issues associated with blind vias but when it comes to glass weaves and differential signals, we end up with a much different animal. This is because the spreading that is done to accommodate blind vias may not be good enough for minimizing skew in differential pairs. And, as noted at the beginning of this article, this has to do with how differential signals operate. When the two members of a differential signal pair travel over the weave pattern, the resulting variations between the two may not be equal in both members and the signal is degraded as a result. Bottom line—addressing weave-induced skew in differential pairs is a tougher nut to crack.

Two of the glass styles or weaves that are known to cause skew in high-speed differential pairs are 106 and 1080. These are pictured in Figure 2 where there the 3.5 mil diameter (89 micron) wire running across the weave represents a trace. To better address the issues associated with skew, a uniform weave style was introduced with 3313 glass. It is shown in Figure 4. Early results showed that skew was held to a minimum using this glass weave. As can be seen, the weave is flat and uniform.

At this point, it would have been great to announce, “Skew problems in differential pairs solved!”

But not every manufacturer fabricates 3313 glass in the same manner. In 2013, at Speeding Edge, we built two different test boards with 3313 glass from two different manufacturers. From the first weaver we got unbelievably good control of skew, while from the second the skew was terrible. In this test, we had 14-inch (35 cm) traces, and for the first weave, the skew was no more than 4 picoseconds. But, for the second weave, the skew was as bad as 62 picoseconds over the same 14 inches. This is 62% of a bit period of 10 Gbps. This equates to a link that is definitely going to fail.

Because there is no standard that specifies the manner in which glass will be mechanically spread, it’s possible to get such disparate results as those we obtained.

Addressing Skew at the Design Stage

Thus far, the methods for addressing differential skew have focused on the mechanical side of the equation—finding a material that has a flat uniform weave and if two different weaves are used in the same type of glass, making sure that the spreading of the weave has been done in the same manner. Additional (or a combination of) efforts include using two plies of glass cloth to avoid voids in the resin bundles, choosing a low DK glass, attempting to align the traces with the pitch of the glass, or orienting the design by 5° on the panel during the PCB fabrication process.

Because dielectric materials are typically woven-glass composites, the Dk varies spatially. As Dk dictates the speed of electromagnetic waves through a medium, variation in Dk means variation in propagation speed of a signal, which is why two signals, even if they travel the same distance, may arrive at different times, based on the particular paths they took. This becomes a critical issue when dealing with very high-speed signals, where the tolerance for skew is a percentage of the unit interval.

 At Avishtech, our ability to address the problem of skew in differential signal lines is directly related to our ability to address the Fiber Weave Effect. As stated above, the weave of the glass fibers in PCB dielectric materials is responsible for the spatial variations in Dk in each dielectric layer. (The dielectric constants of the resin and glass are typically dissimilar because the Dk of the glass is higher than that of the resin). In a plain weave, the percentage of the glass at the knuckles, which are where the warp threads cross over the weft or vice-versa, is highest and, as a result, the Dk is also at a high point. The lowest Dk in a weave would be in open areas, just filled with resin (no glass). Spread glass mitigates this phenomenon by closing the open gaps, but the concentration still won’t be uniform as long as the glass and resin have different Dk values. As noted above, use of materials with low Dk glass, as well as the use of multiple plies can help to mitigate the fiber weave effect but it doesn’t completely take care of skew.

The Fiber Weave Effect Problem and How It Impacts Skew

 As noted earlier, in high frequency, high data rate designs skew budgets become very small, due to the reduced duration of a unit interval, and until now there haven’t been any means to predict or design for skew. It is a strong function of the Dk differences between the glass and neat resin, and the trace widths and spacings relative to the pitch of the glass. However, being a stochastic phenomenon (by nature, due to the alignment of fiber weave and traces), even measuring skew doesn’t provide a meaningful picture, because skew can only really be understood in terms of a distribution.

What the foregoing has meant until now is that product developers were left with some rule-of-thumb skew mitigation techniques and hoping that their final skew numbers didn’t tank their designs.

Stochastic Skew Simulation in Gauss Stack

To address skew in high frequency, high data rate designs, with Gauss Stack, product developers have a very straight-forward, easy way to address skew. The specific steps within the process include:

• Specifying the materials and constructions within the stackup.
• Selecting the copper layer.
• Specifying the trace width/gap and orientation of the trace relative to the warp and weft directions of the glass weave.

Once the foregoing steps have been completed, with a click of a button, the product developer launches a stochastic Monte Carlo stimulation and builds a distribution and confidence levels for the maximum skew on the differential pair. Some applications may still require additional skew mitigation approaches, like the use of rotation and/or re-timers, but the power to predict skew can help the product developer find the solution that either works out the gate or gets as close as possible independent of the use of these additional techniques.

The following example of 3 simulations in Gauss Stack shows how disparate skew distributions can be on the same material.

Figure 5a shows a skew distribution on E-1080 glass for an 85 Ohm differential pair. Because this differential pair has a trace width + gap of 7.5 mils and the pitch of E-1080 glass in the warp direction is ~16.6 mils, this faces a half-pitch problem, where skew becomes extremely high, since the two traces will be nearly completely out of phase with each other (with regard to the periodicity of the glass weave), leading to nearly maximum difference in Dk values seen by the two traces.

Figure 5b shows a skew distribution on E-3313 glass also for an 85 Ohm differential pair. Because the E-3313 glass constructions are thicker, we have a differential pair with wider lines and spacings, taking us a little bit further away from the half-pitch problem we faced with the E-1080 glass, but this still shows skew that is quite high.

Figure 5c shows a skew distribution on the E-3313, except this time the trace width and gap are adjusted to closely match the pitch of the glass fabric (~16.4 mils). By using this full-pitch mitigation technique, we can see that the skew has become nearly zero.

While this confirms that there is merit to the general rules of thumb used for skew mitigation, the disparate scales of impact also makes clear the need to actually be able to quantitatively study this phenomenon and assess its impact on your particular stackup and transmission line designs – something you can now do with Gauss Stack.