Doing jitter timing analysis in the presence of system crosstalk - Embedded.com

Doing jitter timing analysis in the presence of system crosstalk

Serial data standards continue toproliferate, providing dramatic improvements in PC and server systemperformance. Testing these higher speed standards for evidence ofjitter is critical for long-term stability and to achieving theobjective of a good Bit Error Rate (BER) in the design. Effectiveanalysis begins with selecting the right instruments and have a goodunderstanding of instrument noise, rise time and factors such 3rd,4th, 5th harmonic performance.

But it's more than just taking themeasurement — the proper instruments need to be paired with theproper analysis tools. And other factors such as jitter separation,and de-embedding/embedding are also important considerations whentesting serial data rates beyond 8 Gb/sec. For this article, we willfocus on a new approach to jitter separation in the presence ofcrosstalk, a growing problem as the number of lanes increases toboost computing system throughput.

All electrical systems that use voltagetransitions to represent timing information have timing jitter.Historically, electrical systems have lessened the ill effects oftiming jitter (or, simply jitter) by employing relatively lowsignaling rates. As data climb above 8 Gb/sec., jitter has become asignificant percentage of the signaling interval, and understandingthe types and sources of jitter is vital to successfully deployinghigh-speed serial technologies.

At its simplest, jitter is a deviationof an edge from where it should be as shown in Figure 1 . As the ITUdefines it, jitter is “short-term variations of the significantinstants of a digital signal from their ideal positions in time.”

 Jitter is the deviation of an edge's actual occurrance from its expected occurrance

Figure 1. Jitter is the deviation of an edge's actual occurrance from its expected occurrance.

There are several ways in whichjitter can be measured on a single waveform including period jitter,cycle-cycle jitter, and time interval error (TIE), and the designwill often dictate which measurement is appropriate.

In thecase of a stand-alone oscillator, the signal is a clock and it can behopping or swept. Here period jitter is an appropriate measurement.In the case of a transmitter for a serial data stream, the signal isa data stream and ISI (inter-symbol interference) is a key problem. Here TIE jitter is the appropriate measurement.

The engineeron the prowl for jitter issues has a number of instruments available,each with unique strengths and weaknesses:

  • A real-time digital storage oscilloscope (DSO) recovers the whole waveform and can measure anything and can be used for TIE, cycle-to-cycle and period jitter measurements. It has limitations, however, around frequency (or bit rate) and resolution of spectra, minute jitter and multi-level modulation.
  • A BER Tester (BERT) is well suited for TIE jitter, particularly TJ or total jitter, a form of TIE. The advantage of the BERT is that is counts every bit, but the tests can be time consuming to perform.
  • A real-time spectrum analyzer (RTSA) can be used for cycle-to-cycle and period measurements with complex modulations for mobile devices, looking at clocks, PLLs and understanding their dynamic performance.  Limitations include span (sub-100 MHz) and bandwidth signals with large modulation spectrum
  • Equivalent-time sampling oscilloscopes offer the best bandwidth and can be used for all jitter measurements for serial data.  Currently, these are the only instruments with noise analysis and a BER eye. Limitations include no real-time capture and can only be used for repetitive patterns and some jitter spectra are aliased.

One question that often comes up is why worry about jitter ifultimately we're only concerned about the BER.  The reason isthat too much closes the eye (in width) which leads to errors. Jitterand noise analysis are simply tools that let you quickly predict andanalyze problems in the BER. Ultimately, it is all about the errors,but eliminating those errors in a design requires insight into thecause or causes of excessive jitter.

The place to start is togain an understanding of how the system performs from an overall BERperspective. The oscilloscope accomplishes this using eye diagramsand statistical analysis to create a bathtub plot, so named becauseof the shape of the resulting chart as the limits change. With theBERT instrument the result is a jitter peak graph resulting from anexact count of every bit.  As shown in Figure 2 , the jitter peakfrom the BERT on the left and the oscilloscope jitter bathtub plotare nearly an exact equivalent.

 

Equivalent view of BER performance between BERT jitter peak on the left and oscilloscope jitter bathtub on the right. Figure 2. Equivalent view of BER performance between BERT jitter peak on the left and oscilloscope jitter bathtub on the right.

Given the close alignment in results, the oscilloscope is a veryuseful complement to the BERT, since the measurement of TJ to theBER=10-12 can take hours using a BERT and the resultdoesn't reveal what kinds of problems are causing the jitter. The oscilloscope can measure a small amount of data in smart way andthen can break the jitter into jitter components typically followingthe accepted jitter model shown in Figure 3 .

The industry’s jitter model 2001-2010 shows that tital jitter consists of random jitter and deterministic jitter.

Figure 3. The industry's jitter model 2001-2010 shows that tital jitter consists of random jitter and deterministic jitter.

By makingassumptions, the oscilloscope can make TJ@BER calculations thatmirror the results obtained using the BERT in a fraction of the time- that is, if all the assumptions are true.  All models ofcomplex systems make assumptions and simplifications, so the fitbetween the model and the true system behavior will never be exact.As discussed in the remainder of this article, a particularlydaunting problem to date has been crosstalk.

The crosstalk problem

To achieve performance targets, mostserial systems use multiple lanes. As frequencies and data ratesincrease past 10 Gb/s, a small amount of crosstalk can eat up thejitter budget and create timing issues.

Crosstalk occurs when one signal isaffected by a neighboring signal. At high data rates a signalpropagates more like a guided wave than a simple DC current. The waveis guided by the conducting trace but radiates through the dielectricmedium, typically FR4. When more than one signal is present, everyconducting trace on the board includes artifacts of the signals onevery other trace. The accepted terminology is to say that anaggressor signal causes crosstalk on a victim signal. Crosstalkoccurs when the signal of an aggressor is picked up by the conductorguiding the victim signal. Unavoidable discontinuities in circuitlayout, like connectors and vias, where capacitive coupling isgreatest, are critical points that act like antennas in generatingcrosstalk.

Real-time sampling and equivalent-timesampling oscilloscopes use spectrally-based jitter analysistechniques to separate the various jitter components. On real-timesampling equipment, where the frequency components are not aliased,the jitter and voltage noise spectra have sub-harmonic peaks that,rather than appearing as sharp lines, are smeared into broadresonance shapes. On under-sampling equipment, like anequivalent-time sampling oscilloscope, where the spectrum is aliased,crosstalk appears as continuous noise.

In both cases, these spectrally-basedjitter analysis techniques, which measure random jitter (RJ) byintegrating the jitter spectrum continuum, overstate RJ with thecrosstalk timing effects. This leads to an increase in RJ and anoverestimation of TJ. Figure 4 shows oscilloscope measurements ofjitter, in this example a DUT with a large amount of crosstalk.

TJ error in oscilloscopes increases with jitter amplitude as where a BERT has no jitter error. (RTO = real-time oscilloscope, Sampling = equivalent-time oscilloscope).

Figure 4. TJ error in oscilloscopes increases with jitter amplitude as where a BERT has no jitter error. (RTO = real-time oscilloscope, Sampling = equivalent-time oscilloscope).

Crosstalk appears to the oscilloscopeas bounded uncorrelated jitter or BUJ since it follows a boundeddistribution. The bounded nature of the distribution is obscured bythe complexity of the data pattern. The seemingly random distributionof 1s and 0s causes different amounts of voltage noise to betransmitted on each aggressor-signal transition.

The vulnerability to crosstalk-inducedBUJ differs between measurement systems. Oscilloscope measurements orextrapolations of jitter pessimistically bundle BUJ or NP-BUJ intoRJ, and then over-report TJ as well. Jitter results (RJ, TJ) dependstrongly on aggressor pattern complexity, with PRBS31 being theworst. PRB7, on the other hand, typically does not cause a largeerror. In the case of real-time oscilloscopes, RJ and TJ results alsodepend on record length, and longer record length provides moresample points to depict better separation. The exact mechanism of theproblem is also implementation dependent.

BUJ measurement solutions

Currently there are a number ofapproaches to jitter analysis on signals where crosstalk issuspected, but none of them provide one-button push results similarto what oscilloscopes can provide for DDJ and PJ. One clue is if thejitter analyzer reports an inordinately large RJ measurement. It israre that thermal effects, the ultimate cause of RJ, manage toconspire to greater than 3 ps RMS. If the RJ reported is larger than3 ps then it's likely that crosstalk is causing problems.

Other tricks to identifying crosstalkrequire more control over the aggressor channel. For example, if it'spossible to turn off the suspected aggressor signal, then you cancompare the RJ measurement with and without a signal on theaggressor. If RJ-with aggressor > RJ-without then the problem iscrosstalk. A work-around is to use the measurement of RJ with theaggressor off and the measurement of dual-Dirac DJ with the aggressoron in the Dual-Dirac model to estimate the Total Jitter of the systemat the BER of interest. The problem with this approach is that itrequires control of the aggressors which is not always possible.Another issue is that it is invalid in non-linear systems (which mosttransmitters are), and is optimistic toward errors since some of thecrosstalk is unbounded.

A more advanced approach would be to beto implement BUJ-aware jitter analysis algorithms. These wouldinvolve an additional step in jitter analysis after separation of DDJand PJ to separate NP-BUJ from RJ as shown in Figure 5 . A keyadvantage is that this will work in every scenario since no controlover the aggressor is necessary and non-linear TX does not present aproblem. Further, unbound crosstalk components would be correctlyrecognized as unbound. The downside to this approach is somepessimism remains in the result.

Bounded uncorrolated jitter is now a component of deterministic jitter.

Figure 5. Bounded uncorrolated jitter is now a component of deterministic jitter.

To test the ability of a jitteranalysis algorithm to accurately separate BUJ from other randomjitter sources, we repeated the test shown in Fig. 4, but with theaddition of the results from an equivalent-time sampling oscilloscopeusing a BUJ-aware jitter analysis algorithm. The result indicated bythe dashed line in Figure 6 still shows some pessimism compared tothe BERT. Results with a real-time oscilloscope were slightly morepessimistic. That said reported TJ error accuracy is dramaticallyimproved, making it possible to trust oscilloscope TJ measurementseven in designs where you suspect crosstalk may be a source of jitterand noise related errors.

 

BUJ-aware jitter analysis (dashed line) algorithm shows greatly improved accuracy on a DUT with large crosstalk.

Figure 6. BUJ-aware jitter analysis (dashed line) algorithm shows greatly improved accuracy on a DUT with large crosstalk.


Summary

As data rates continue to increase,jitter has become a significant percentage of the signaling interval,making it increasingly important for designers to fully understandthe types and sources of jitter in their designs. Since mosthigh-speed serial design now involve multiple lanes, crosstalk is anearly unavoidable consequence that must be factored into the jitterbudget.

But to date, measuring the effects ofcrosstalk-induced jitter, or bounded uncorrelated jitter has beennotoriously difficult using jitter separation techniques. BecauseBUJ has not been accounted for in the jitter algorithms it has beenlumped with RJ leading to pessimistic total jitter results comparedto the result obtained from a BER tester.

In recognition of this growing problem,particularly for data rates above 10 Gb/s, the jitter model isexpanding to include BUJ with the addition of BUJ-aware algorithms.In tests involving a large amount of crosstalk, the new models haveproven effective at delivering TJ results on real-time andequivalent-time sampling oscilloscopes consistent with those from aBERT. It also allows for a more thorough analysis of jitter problemsin a design, including jitter induced by crosstalk.

Chris Loberg is a seniortechnical marketing manager at Tektronix responsible foroscilloscopes in the Americas region. Chris has held variouspositions with Tektronix during his 13 years with the company,including marketing manager for Tektronix' Optical Business Unit. Hisextensive background in technology marketing includes positions withGrass Valley Group and IBM. He holds an MBA in marketing from SanJose State University. His article was was also published on Embedded.com's sister publication, EDN Network.  

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.