Finding Sources of Jitter with Real-Time Jitter Analysis

Application Note 1448-2

As data rates increase effects of jitter becomes critical andjitter budgets get tighter.Instruments such as Bit ErrorRatio Testers (BERT) areoptimized for determining thetotal amount of jitter andworst-case eye-opening in yourhigh-speed digital system and canbe used to test for compliancebased on industry standards. In addition, some real-timeinstruments can separate random and deterministic jittercomponents to predict/extrapolateworst-case total jitter (TJ) and eye-opening based on auser-specified Bit Error Ratio(BER), typically 10-12. But whenjitter measurements do not meeta particular minimum standard,or if jitter measurement resultsare “too close for comfort,” then measuring the amount of component or system jitter is just half of the jitter testequation. Determining theroot-cause of jitter is the otherhalf of the test equation. Thefocus of this paper will be to

address some practical “tips & tricks” on using real-timeoscilloscopes with jitter analysisand high-speed pulse/patterngenerators to separate andtime-correlate specificdeterministic jitter componentsto help identify, measure, andview sources of systematic timing errors.

Understanding Jitter

Jitter is the deviation of a timingevent of a signal from its idealposition. The traditional way to measure jitter is with aneye-diagram using repetitive

acquisitions on an oscilloscope as shown in Figure 1. Looking atthis composite view, you mightassume that you have a band ofworst-case jitter equal to thewidth of the rising and fallingedges of the eye-diagram. Youmight also assume that all edgesof your signal are jittering aboutto the same degree. Both of theseassumptions would be incorrect.

Jitter is complex and is composedof both random and deterministicjitter components. Random Jitter(RJ) is theoretically unboundedand Gaussian in distribution.“Unbounded” simply means that

Figure 1. Viewing jitter in an eye-diagram.

2

if you wait long enough, thepeak-to-peak jitter will increaseindefinitely theoretically. Thismeans that an eye-diagram maynever show the worst-casecondition. If the jitter in yoursystem consisted of just randomjitter, then each edge of the datasignal would have the sameprobability of timing error.

Deterministic Jitter (DJ) isbounded and doesn’t follow anypredictable distribution. Itconsists of several differentsub-components and is usuallycaused by a systematic problemin your high-speed digital design.For this reason, DJ is sometimesreferred to as systematic jitter. Ifyou were to view each individualedge of the data signal, you wouldprobably see that particular dataedges contribute different amountsof timing error. Depending on thedata pattern, some edges in theserial pattern will always beshifted to the right (positivetiming error), while other edgeswill always be shifted to the left(negative timing error) relative totheir ideal timing locations. Andthen the random jitter componentwould cause these individual databit edges to randomly bouncearound the shifted/offsetdeterministic amplitude of jitter.

Using an eye-diagram, you cansometimes quickly determine ifyour system is dominated byrandom or deterministic jitter, or

possibility a combination of thetwo. Using the oscilloscope’svariable intensity or color-gradeddisplay capability, you canvisually look for the existence of “bright” trace paths within the infinite-persistence display.Referring to the traces in Figure 1,we can see several well-definedpaths of brightness. This is aclear indication of deterministicjitter. Some of the data edgesconsistently occur in somelocations while other edges occur in other locations of theeye-diagram. The power of areal-time scope is the ability toview these particular data edgesindividually as we will see later in this paper.

Another technique to determinethe existence of random versusdeterministic jitter is to use thescope’s histogram feature. Bytaking a look at a slice of dataacross the center of the screen,we can see in Figure 1 that thedistribution of edge placements is a combination of random and deterministic jitter. If theProbability Distribution Function(PDF) were Gaussian (the classicbell-shaped curve), then thesystem jitter would probably bedominated by random jitter. Thefact that this particular PDF is approximately bi-modal indistribution is another indicationof the existence of a significantamount of deterministic jitter.

Real-Time Jitter Analysis

Measuring the total jitter in yoursystem with a Bit Error RatioTester (BERT) or real-time RJ/DJseparation techniques will tellyou whether or not your systemmeets a jitter/timing budgetspecification. And viewing aneye-diagram along with ahistogram can give you a goodintuitive feel concerning the typeof jitter in your system and arough feel for the amount of totaljitter. But neither of these twotypes of measurements canprovide you with much insightinto how to identify, view andthen reduce specific componentsof jitter. This is where real-timejitter analysis steps in. Theprimary contribution of jitteranalysis using a real-timeoscilloscope is its inherent abilityto capture data or clock pulses ina single acquisition with timingmeasurements on each and everypulse in a long stream of data.The real-time scope and jitteranalysis can then be used totime-correlate specific jitter error measurements to specificdata bits or other signals in yoursystem that might be contributingto the total system jitter.

3

Data

Clock

Trend

Spectrum

Histogram

Figure 2. Real-time jitter analysis.

Figure 2 illustrates the techniquethat real-time jitter analysis usesto measure jitter on an NRZ datasignal. This type of real-time jittermeasurement is typically referredto as a Time Interval Error (TIE)or phase jitter measurement. Thereal-time scope first captures adeep record of the NRZ datasignal and then creates an idealsoftware-recovered clock basedon the captured data. Dependingon the user’s selection, thissoftware-generated clock can beeither of a fixed frequency or aPLL-type clock with a specifiedloop bandwidth. The jittersoftware then performs a best-fitalgorithm to align the NRZ dataedges with the “virtual” clockedges. Although Figure 2 shows a software-generated clockwaveform (blue trace), this signalis typically NOT shown on thescope’s display. It is merely usedas a theoretical timing referencefor making Time Interval Error(TIE) measurements on each dataedge by the scope’s computer.

Jitter analysis software in thereal-time oscilloscope thenperforms a series of delta-time

measurements between each edgetransition in the data record at aspecified threshold level relativeto edges of the best-fit clock.There is no need to specify ameasurement threshold level for the clock since it is just atheoretical time reference. Aftercompleting all of the timingmeasurements in the data record,the results can be viewed in threedifferent formats.

With the jitter “trend” waveform,absolute timing errors for eachdata edge are plotted on thevertical axis with the horizontalaxis based on the scope’stimebase. In other words, time error versus time. The“trend” waveform provides atime-correlated view of eachtiming error relative to the data waveform. In the case ofdata-dependent jitter, the “trend”waveform can be a very powerfultool to time-correlate specificerrors with the data bit thatcaused the error. In the case ofPeriodic Jitter (PJ) caused bysignal coupling, the “trend”waveform can also betime-correlated to other signals

in the system captured on otherchannels of the oscilloscope.Characteristics of periodic jitterwill be discussed later in this paper.

Another way to view the jitter isin the frequency domain. Withthe jitter “spectrum” waveform,an FFT math function isperformed on the TIE trend datato produce a view of the jitterbased on repetitive frequencycomponents within the series ofdelta time error measurements.In this case, data is plotted asamplitude timing error on thevertical axis versus frequency onthe horizontal axis. This view canbe especially beneficial whenlooking for uncorrelated PeriodicJitter (PJ) components that arenot time-correlated to the data signal.

The “histogram” view will showthe Probability DistributionFunction (PDF) of the jitter (composite of all TIEmeasurements in the data record)and is plotted as timing errorversus number of hits (N). Theresults of the real-time histogramshould closely correlate with theresults of a repetitive histogramproduced from an eye-diagrammeasurement. However, withreal-time sampling much moredata is collected from a singleacquisition of the signal. In otherwords, you will be able to view adistribution of jitter from a singleacquisition. In addition, thereal-time jitter histogram willbuild up with repetitive real-timeacquisitions to generate a moreaccurate and complete PDF.

4

Data

Clock

Trend

Figure 4. Duty Cycle Distortion (DCD).

TJ

RJDJ

DCD ISI PJ

Figure 3. Jitter components.

Typical Characteristics ofIndividual Jitter Components

In order to interpret measurementresults and waveform viewsperformed by real-time jitteranalysis, you must firstunderstand the characteristicsand likely causes of individualjitter components. There is more to jitter interpretation than just knowing that RandomJitter (RJ) is Gaussian indistribution, and deterministicjitter is non-Gaussian.

As mentioned earlier, Total Jitter(TJ) is composed of a RandomJitter (RJ) component and aDeterministic Jitter (DJ)component. Random jitter isunbounded, and for this reason(unlimited peak-to-peak) RJ isusually measured in terms of anRMS value. In addition, randomjitter is very predictable in termsof distribution. The ProbabilityDistribution Function (PDF) isalways Gaussian in distribution.Unfortunately, predicting thecause of RJ is a more difficult taskand is not within the scope of

this paper. RJ is often caused bythermal effects of semiconductorsand requires a deeper understandof physics. However, one piece ofadvice is to pay close attention tothe amount of vertical noise inyour system. Random verticalnoise will directly translate intorandom timing jitter.

On the other hand, deterministicjitter is bounded and is alwaysmeasured in terms of apeak-to-peak value. Although thedistribution of deterministic jittercan be very unpredictable, thelikely causes and characteristicsof the individual sub-componentsof measured deterministic jitter are very predictable. Thesub-components of DeterministicJitter (DJ) consist of Duty CycleDistortion (DCD), Inter-SymbolInterference (ISI), and PeriodicJitter (DJ) as shown in Figure 3.Let’s now take a closer look atpossible causes and characteristicsof each of the sub-components ofdeterministic jitter.

There are two primary causes of Duty Cycle Distortion (DCD)jitter. If the data input to a

transmitter is theoreticallyperfect, but if the transmitterthreshold is offset from its ideallevel, then the output of thetransmitter will have duty cycledistortion as a function of theslew rate of the data signal’s edgetransitions. Referring to Figure 4,the waveform represented by thedotted line shows the ideal outputof a transmitter with an accuratethreshold level set at 50% andwith a duty cycle of 50%. Thesolid line waveform represents adistorted output of a transmitterdue to a positive shift in thethreshold level. With a positiveshift in threshold level, theresultant output signal of thetransmitter will have less than50% duty cycle. If the thresholdlevel is shifted negatively, thenthe output of the transmitter willhave greater than 50% duty cycle.

Measuring TIE relative to thesoftware-generated best-fit clockresults in a positive timing erroron the rising edge of each data bitand a negative timing error onthe falling edge of each data bit.The resultant TIE trend waveformwill possess a fundamental

5

frequency equal to 1/2 the datarate. The phase of the TIE trendwaveform relative to the datasignal will depend on whether the threshold shift is positive ornegative. With no other sources of jitter in the system, thepeak-to-peak amplitude of DCD jitter will be theoreticallyconstant across the entire datasignal. Unfortunately, othersources of jitter, such as ISI,almost always exist making itsometimes difficult to isolate the DCD component. But onetechnique to test for DCD is tostimulate your system/componentswith a repeating 1-0-1-0… data pattern. This technique will eliminate Inter-SymbolInterference (ISI) jitter and makeviewing the DCD within both thetrend and spectrum waveformdisplays much easier. Using thejitter spectrum display, the DCDcomponent of jitter will show upas a frequency “spur” equal to 1/2the data rate.

Another cause of DCD isasymmetry in rising and fallingedge speeds. A slower falling edgespeed relative to the rising edgewill result in greater than 50%duty cycle for a repeating1-0-1-0… pattern, and slowerrising edge speeds relative to thefalling edge will result in lessthan 50% duty cycle. Although notgraphically shown in this paper,the results of jitter analysis andthe TIE trend waveform will looksimilar to the results of theexample illustrated in Figure 4.

Inter-Symbol Interference (ISI),sometimes called data dependentjitter, is usually the result of abandwidth limitation problem ineither the transmitter or physicalmedia. With a reduction intransmitter or media bandwidth,limited rise and fall times of thesignal will result in varyingamplitudes of data bits dependenton not only repeating-bit lengths,but also dependent on precedingbit states. In addition, improperimpedance termination andphysical media discontinuitieswill also result in ISI due toreflections that cause signaldistortions. Although we willaddress these two phenomena(BW limitations and reflections)separately in this paper ascontributors to ISI jitter, inreality, waveform distortions due to reflections are also abandwidth limitation problem.

Figure 5 shows an example of ISI due to bandwidth limitationproblems. Limited bandwidthproduces limited edge speeds,and limited edge speeds willresult in varying pulse amplitudesat high-speed data rates. Varyingpulse amplitudes will then resultin transition timing errors. Let’snow examine this more closely.

With a long series of repeating“1’s,” the amplitude of the datasignal will eventually rise to a full steady-state high level asillustrated by the long-high pulseat point A in Figure 5. When thestate of the data changes to a “0,”the signal will have a relativelylong transition time to reach thethreshold level, resulting in apositive timing error. This will bemanifested as a positive peak oftiming error in the jitter trendwaveform at point B. Note thatthis point on the jitter trendwaveform is time-aligned with thenegative data crossing point onthe data signal.

A

B

C

D

Figure 5. Inter-Symbol Interference (ISI) due to BW problem.

6

A

B

Figure 6. ISI due to reflections.

The negative peak amplitude of the next “0” bit preceded by a long string of “1’s” will beattenuated for two reasons. Firstof all, the preceding long string of“1’s” means that the signal willtake longer to transition to a truelow level since the data signalstarts from a higher initial level.Secondly, the following “1” bitcauses the signal to reversedirection before it even reaches asolid low level. This reduction insignal amplitude will produce anegative timing error on the nexttransition to a “1” since the signalhas a very short distant to travelto reach the threshold level. Thisis illustrated at point “C” on thejitter trend waveform.

The positive timing errorillustrated at point “D” on thejitter trend waveform follows thesame logic as the positive timingerror at point “B” previouslydiscussed. With a long string of“0’s” the data signal has sufficienttime to settle to a full steady-statelow level. When this signal thentransitions back to a high level, itagain has a longer transition timeto reach the threshold level, andhence produces a positive timing error.

Once you understand howbandwidth limitations produceInter-Symbol Interference (ISI)timing errors, it becomes moreintuitive to understand theunique “signature” of the jittertrend waveform due to ISI and how it relates to the

time-correlated serial data signalunder measurement.

Another common cause ofinter-symbol interference besidesbandwidth limitations is signalreflections due to improperterminations or impedanceanomalies within the physicalmedia. Signal reflections willproduce distortions in theamplitude of the data signal asshown in Figure 6. Depending on physical distances betweenimpedance anomalies, reflectionsproduced by one pulse may notshow up on a high-speed datasignal until several bits later inthe serial pattern. Notice whichpulse the arrows begin on and the where pulse distortion(reflection) occurs as illustratedby the end of each of the arrowsin Figure 6.

If the amplitude of the signalbecomes distorted on or near

a data transition edge due toreflections, then a timing errormay occur. If a signal reflectioncauses signal attenuation nearthe data edge, then a negativetiming error will be detectedsince the signal will have less distance to travel whentransitioning to the thresholdlevel. This is illustrated at point“A” on the jitter trend waveformin Figure 6. If a signal reflectioncauses a boost in signalamplitude, then the result will be a positive timing error sincethe signal will have farther totransition to reach the thresholdlevel. This is illustrated at point“B” on the jitter trend waveform.Inter-symbol interference due tosignal reflections can be verydifficult to isolate and interpret.But if you have signal reflectionproblems in your system, youprobably also have a bandwidthlimitation problem.

7

Periodic Jitter (PJ) is usually theresult of a cross-coupling or EMIproblem in your system and can be either correlated oruncorrelated to the data signal.An example of uncorrelated PJwould be signals from a switchingpower supply coupling into thedata or system clock signals. It is considered to be uncorrelatedbecause it is not time-correlatedwith either the clock or datasignal since it would be based on a different clock source. Anexample of correlated PJ wouldbe coupling from an adjacent datasignal based on the same clock ora clock of the same frequency.

Figure 7 shows an example of a“corrupter” signal (upper trace)capacitively coupling into ourserial data signal (middle trace).This coupling will result inamplitude distortions on the data signal. Just like inter-symbolinterference due to reflections, if these amplitude distortionsoccur at or near a data signaltransition, a timing error may occur.

Since most Periodic Jitter (PJ)will be uncorrelated to the data signal, any attempts totime-correlate the jitter trendwaveform to the data waveformmay turn into a futile attempt. Aswe will show later in this paper,uncorrelated periodic jitter canoften be detected using the jitterspectrum view.

Figure 7. Periodic Jitter (PJ) caused by capacitive coupling.

8

Isolating Jitter Components in the Real World

In the examples that we have justillustrated, we have shown howindividual and isolated jittercomponents are theoreticallymanifested in the jitter trend andspectrum views. But in the realworld, jitter components arerarely isolated. If you havemultiple jitter components inyour system contributing to thetotal system jitter, these variousjitter components becomeconvoluted and you end upviewing composite results, whichcan be difficult to interpret. Ifyou are familiar with the arcadegame called “whack-a-mole,”perhaps you can relate to thisanalogy. In this arcade game the“moles” pop their heads up one at a time, and it’s your job to“whack” them on their heads witha mallet to push them back down.This is exactly what we would like to do with the various jittercomponents. It would be nice

if the jitter “moles” (jittercomponents) would appearindividually so that we couldidentify them and then “whack”them down. Unfortunately, in alive system all of the jitter molesmay have their heads popped upsimultaneously. This makes itdifficult to identify their sourceand to decide which “mole” towhack first.

But there are some novelstimulus-response techniques you can employ in order toisolate, measure, and then viewindividual jitter components.Once you successfully isolateindividual components, you can then often time-correlateworst-case peaks of jitter tospecific data bit transitions andthen use common-sense debugtechniques to solve your jitterproblems… whacking them downone component/mole at a time.

The primary tool to isolate jitter components is a real-timeoscilloscope with responsive andinteractive jitter analysis such asAgilent’s 6 GHz 54855A, shown inFigure 8. In addition, a high-speedpulse/pattern generator such asAgilent’s 3.35 Gb/s 81134A, shownin Figure 9 can be very useful forgenerating known serial patternsof high-speed differential stimulus.Let’s begin with some real jittermeasurement examples usingthese two measurement tools.

Figure 8. Agilent’s 54855A 6 GHz Real-time Oscilloscope.

Figure 9. Agilent’s 81134A 3.35 Gb/s Pulse/Pattern Generator.

9

Isolating and Measuring DutyCycle Distortion (DCD)

One technique mentioned earlierto isolate and measure Duty CycleDistortion (DCD) is to stimulateyour system/component with a repeating 1-0-1-0… serialpattern. This stimulus patternwill eliminate most of theInter-Symbol Interference (ISI).Although ISI is eliminated withthis repeating pattern, RandomJitter (RJ) and any Periodic Jitter(PJ) will still be present in thesignal, which will contribute toconvoluted measurement results.But there is a measurementtechnique to also eliminate bothrandom jitter and uncorrelatedperiodic jitter in the jittermeasurement results. Figure 10shows the capture of a repeating1-0-1-0… serial pattern (greentrace (top)) with jitter analysisresults showing the TIE trendwaveform (orange trace(bottom)). To eliminate therandom components (RJ and PJ),we have used waveform math toaverage the jitter trend waveform. Before averaging, the TIE trendwaveform would be “bouncing”vertically due to the randomcomponents with repetitiveacquisitions. But averaging haseliminated the random jittercomponents to show a very stable trend waveform with anapproximate constant level ofpeak-to-peak amplitude of jitterfrom cycle-to-cycle. We can thenuse the scope’s manual markersor the scope’s automaticparametric measurementcapability to measure thepeak-to-peak amplitude of dutycycle distortion. In this case, wemeasured approximately 10 ps of DCD.

In addition to determining thelevel of DCD, we can also gleanadditional information about ourmeasurement results. With thetime-correlated display of thejitter trend waveform and datasignal, we can see that the trendwaveform is in-phase with thedata signal. This is an indicationthat the duty cycle of our pulse is less than 50%. Rising edgesalways occur late (+error), andfalling edges always occur early(-error). Perhaps our transmitterthreshold level is too high, orperhaps the output of the ourtransmitter generates slowerrising edge speeds as compared tofaster falling edge speeds. At thispoint if we believe thepeak-to-peak level of DCD isexcessive, we can setupadditional characterization teststo measure the duty cycle of eachpulse in the data stream usingjitter analysis. In addition, if wesuspect that the duty cycle

distortion is caused byasymmetry in the rising andfalling edge speeds, we can alsosetup the instrument and jitteranalysis to characterize eachrising and falling edge in the data stream.

One critical element to performthis particular measurementusing waveform averaging is that the oscilloscope’swaveform-capture and jitteranalysis capability must be veryresponsive with a fast display andmeasurement update rate. Somejitter analysis packages aregeared more for single-shot/staticmeasurements. Averagingrequires multiple acquisitions ata fast rate. Agilent’s MegaZoomtechnology provides the fastestdisplay update rates andmeasurement processing time on deep memory records in theoscilloscope industry.

Figure 10. Isolating DCD.

10

Isolating and Measuring ISI

Figure 11 shows a measurementexample of isolating inter-symbolinterference due to a systembandwidth limitation problem.Using a high-speed pulse/patterngenerator, stimulate your selected devices under test with a repeating PRBS signal at thefull-specified clock rate. You mayapply the stimulus at either theinput of a transmitter circuit, theinput of a receiver, or possibly atthe input of a transmission line infront of a receiver. Using one ofthe scope’s channels, capture thedata signal at the desired pointand setup jitter analysis toperform a TIE measurement withthe jitter trend waveform view.Running real-time acquisitionsrepetitively with a synchronoustrigger, you will observe the TIEtrend waveform (orange trace(bottom)) randomly bouncing upand down in amplitude. This“bouncing” is primarily due toRandom Jitter (RJ). But if thescope’s update rate is sufficientlyfast, you can usually see arepetitive pattern of thedeterministic components. Toeliminate the random components,simply use waveform math toaverage the TIE trend waveform.With sufficient cycles ofacquisition and averaging, theresultant TIE trend waveform willbecome very stable and consist of just the Deterministic Jitter(DJ) components.

Once you have a stable jittertrend waveform, you can thentime-correlate specific peaks ofjitter back to the serial datasignal. The TIE trend waveform

will have a unique “signature” of DJ dependent on the serialdata pattern. If yoursystem/components exhibitbandwidth limitation problems,you can observe and time-correlateinter-symbol interference jitter asevidenced by positive peaks ofjitter coincident with the end oflong strings of “1’s” or “0’s.” And you will probably observenegative peaks of jitter coincidentwith the end of short “1’s” or “0’s.”

To measure the amount ofdeterministic jitter, simply select to perform an automaticpeak-to-peak measurement on theaveraged TIE trend waveform.With minimal correlated PeriodicJitter (PJ) in the system, thismeasurement gives the totalpeak-to-peak deterministic jitterconsisting of just Duty CycleDistortion (DCD) and Inter-SymbolInterference (ISI) jitter

components. In this example, we measured approximately 44 psof deterministic jitter. From theprevious measurement where we determined DCD using arepeating 1-0-1-0… pattern, wecan compute the approximate ISIcomponent as 44 ps – 10 ps = 34 ps.

In addition to measurementresponsiveness, another criticalelement of being able to performjitter measurements such as theseon high-speed serial data signalsis having an oscilloscope anddifferential active probes withsufficient bandwidth. Poorlydesigned probes or long cablescan contribute a significantamount of inter-symbolinterference jitter to yourmeasurements. You definitelydon’t want to wind up measuringjitter components caused bywaveform distortions contributedby your scope, probes, and cabling.

Figure 11. Isolating Inter-symbol Interference (ISI).

11

Figure 12 shows an example ofviewing inter-symbol interferencegenerated by both bandwidthlimitation problems andreflections. It is almost impossibleto separate termination problemsfrom bandwidth limitationproblems when making jittermeasurements. Both of thesecontributors will result in ISIjitter for the same reason, whichis modulation of peak amplitudesof pulses at the receiver. Thesemodulations will ultimatelygenerate timing errors because ofvariations in transition times todata crossing threshold levels.But there are some clues that you can look for to identifyimpedance termination problemsthat result in signal reflections.

If termination problems do notexist, then leading-edge pulseshapes of similar width pulsesshould look similar since eachpulse is generated by the sametransmitter and travels down the same physical media. But ifreflections do occur, you willprobably be able to see dissimilarpulse shapes. In Figure 12 we show an example of twohigh-level pulses (A & B)consisting of a series of “1’s.” Butnotice the differences in theirpulse shapes near the leadingedges. The first long series of“1’s” (A) is not immediatelypreceded by other pulses. But we can see a reflection (an “up”bump) approximately 1 ns after

the leading edge of this pulse. Thesecond long series of “1’s” (B) isimmediately preceded by twoshort “1’s.” We can see distortionin this pulse (another “up” bump)approximately 1 ns after thepreceding short “1” pulse. Due to termination problems, thispreceding pulse is causing adelayed reflection onto this laterpulse very close to its leadingedge. Depending on the datapattern, reflections will also bepresent on some of the shorterpulses, but these are much moredifficult to identify becausereflections on narrow pulses willbe “masked” due to bandwidthlimitations on short “1” and “0” pulses.

Measuring the total effect of ISI jitter including terminationproblems is exactly the same asjust previously discussed. Simplyestablish a synchronous triggerand then average out the randomjitter components from the TIEtrend waveform. The net result is the total Deterministic Jitter(DJ), which consists of acombination of ISI and DCD. For this example we measuredapproximately 47 ps ofpeak-to-peak deterministic jitter.Then simply subtract out the DCD component based the DCDmeasurement using the repeating1-0-1-0… serial pattern. If your ISImeasurement results appear to be excessive, you can then easilytime-correlate worst-case jitterpeaks with specific data pulses.

Figure 12. ISI due to reflections.

A B

12

Network Analyzer (VNA) to more accurately characterize your physical media forimpedances and distancesbetween physical anomalies.

Isolating Uncorrelated PeriodicJitter (PJ)

Periodic jitter (PJ) is typicallycaused by cross-talk, signalcoupling, or EMI. In most cases,PJ will be uncorrelated to thedata signal, meaning the“corrupter” signal has no timerelationship to the serial datasignal. The best tool to use to identify the presence of

uncorrelated PJ is usually the jitter spectrum display.Uncorrelated PJ will exhibitunique frequency “spurs” in thespectrum display. Within thespectrum display you willprobably see lots of spurs. Most of these spurs will be harmonicsand sub-harmonics of the datasignal, which indicates thepresence of pattern dependentjitter such as DCD and ISI. Whentesting for uncorrelated periodicjitter, it will be your job to lookfor “odd ball” frequency spursunrelated to the data signal frequency.

To verify that the waveformdistortions you are viewing onthe data signal are really causedby reflections and not by someother effect, such as probingresonance, probe your signal at the transmitter side of thetransmission line and reduce the data rate of your signal. Asyou can see in Figure 13, thereflections will become morepredominant at the transmitterside. And with a slower data rate,it will be easier to determinewhich pulse edge causes eachwaveform distortion and then youcan easily measure the time delaybetween the cause and effect tohelp identify the location of yourimpedance discontinuities. In thisparticular example, we haveidentified that our transmissionline has both a capacitive andinductive discontinuity asevidenced by the positive peakreflections (low impedance) andthe negative peak reflections(high impedance) respectively.

Measuring reflections at the transmitter side of thetransmission line with a reduced data rate signal is very similar to a Time DomainReflectometry (TDR)measurement. In fact, if yourreflection problems are seriousenough, it might be advisable touse a real TDR instrument suchas Agilent’s 86100A, or a Vector

Figure 13. Testing for reflections at the transmitter side.

13

Referring to Figure 14, we havecaptured a long string of serialdata (yellow trace (top)) usingdeep memory and then performedjitter analysis on each data edge. The purple trace (middlewaveform) shows the jitter trendwaveform while the orange trace(bottom waveform) shows thefrequency spectrum of jitter.Looking at the TIE jitter trendwaveform, it is impossible isvisually detect any low frequencymodulation in this measurement.However, the jitter spectrum view(orange trace (bottom)) clearlyshows a low frequency componentof modulation. In this example,we have measured a “spur” on the spectrum display at exactly400 kHz. This frequency isunrelated to the data signal’s2.5 Gb/s data rate and is anindication that we have justdiscovered an uncorrelatedperiodic jitter component.

Note that if the periodic jitter issinusoidal in nature, then thespectrum display will generate asingle frequency spur associatedwith the fundamental frequencyof this coupling component.However, if the periodic jitter isnon-sinusoidal, then there willprobably be several frequencyspurs associated with thiscomponent of coupling. But all of these particular spurs on the spectrum display will beharmonics of one another. Tohelp identify the source of digitalperiodic jitter, look for thefundamental frequency componentamongst the cluster of spurs.

Just knowing the frequency of the modulation should give us aclue as to its source. In this case,we suspect that our system’sswitching power supply may be

Figure 14. Isolating uncorrelated PJ with a jitter spectrum measurement.

Figure 15. Verifying that switching power supply coupling causesperiodic jitter.

coupling into our data signal,perhaps via the power supplies.To verify this suspicion, we have used another channel ofoscilloscope to capture ripple on a power supply that is

synchronous with the switchingsignal as shown by the greentrace (second from top) inFigure 15. We set up the scope totrigger on this suspect corrupterand then used a smoothing

14

this measurement by capturing a minimum of one period ofuncorrelated PJ. However, withdeeper memory acquisition youcan obtain higher resolutionmeasurements in the spectrumdisplay. In addition, deepermemory will enable you tomeasure lower frequencycomponents of jitter. With the54850 Series scopes 1 M points of acquisition memory, you candetect jitter components as low as 20 kHz.

Isolating Correlated Periodic Jitter (PJ)

Detecting the presence ofcorrelated Periodic Jitter (PJ) isusually the most difficult type ofjitter to uncover. Fortunately, thistype of jitter is usually the leastlikely source of jitter contributingto a system’s total jitter. The fact that it is correlated meansthat this jitter component issynchronous with the serial datastream that it is corrupting. A

possible source of correlated PJ iscoupling from an adjacent lane ofserial data, such as multi-lanedInfiniBand data traffic. In thiscase, adjacent lanes of data wouldbe based on a common clock.

Since correlated periodic jitter is time-correlated to the serialdata signal under test, it may bedifficult to detect this componentof jitter using the jitter spectrumdisplay. Correlated PJ willdefinitely contribute to specificfrequency spurs in the spectrumdisplay, but these spurs may landon top of or between other datadependent spurs making it hardto discriminate between ISI and correlated PJ frequencycomponents in the jitter spectrum display.

One technique that may provesuccessful for you is to set upyour system to generate serialdata traffic on just one data laneas shown in Figure 16. Using thescope’s deep memory, engage TIE

function on the jitter trendwaveform (purple trace (thirdfrom top)). With an appropriatesmoothing factor, the smoothingfunction will eliminate the higherfrequency components of jitter and“draw-out” the lower-frequency,uncorrelated periodic jittercomponent in the jitter trendwaveform. If the smoothed trendwaveform (purple trace (thirdfrom top)) appears to be lockedonto and synchronous with thetrigger source signal (green trace(second from top)), then we havepositively identified the source of corruption.

Using either the scope’sautomatic measurements ormarkers, we can now easilymeasure the peak-to-peakamplitude of this jitter component.In this example, we measureapproximately 5 ps peak-to-peakof uncorrelated PJ caused bycoupling from the switchingpower supply, which is aninsignificant amount of jitter. But if the amplitude of this jittercomponent is excessive, thenperhaps better shielding,filtering, or improved board tracelayout techniques are needed to reduce the amount of thiscomponent of periodic jitter in our system.

Having sufficient acquisitionmemory depth in youroscilloscope is a criticalmeasurement element to detectrelatively low frequency periodicjitter. In this example, we havecaptured eight periods of lowfrequency modulation using 400 kpoints of acquisition memorywith the A-to-D converterrunning at its maximum samplerate of 20 GSa/s. A scope withless memory could have made

Figure 16. Detecting correlated Periodic Jitter (PJ).

15

and then perform the same DJmeasurement with the systemgenerating multi-lanes of datatraffic. The peak-to-peakdifference between these twomeasurements is the amount ofpeak-to-peak correlated PJ inyour system. To compute the total amount of PJ (correlatedand uncorrelated), simply addboth PJ components.

Isolating Random Jitter (RJ)

In several of the previousexamples of testing forinter-symbol interference, we have shown how you caneliminate random components of jitter (RJ + uncorrelated PJ)from the jitter trend waveform by establishing a synchronous

trigger on the data signal andthen averaging the jitter trendwaveform using waveform mathand repetitive acquisitions. Toisolate the random components,simply use a chained waveformmath function as shown inFigure 17 to subtract the averagedtrend waveform (green trace(second from bottom)), whichrepresents the DJ component,from the real-time trendwaveform (purple trace (thirdfrom bottom)), which representsTJ. The results will be a trendwaveform (pink trace (bottom))consisting of both Random Jitter(RJ) and uncorrelated periodicjitter. Since these components are random relative to the datasignal, it will be impossible totime correlate these results to thedata waveform.

Figure 17. Measuring system Random Jitter (RJ).

jitter analysis with the spectrumdisplay turned on and average theresults using waveform math andrepetitive acquisitions (bottomtrace). Averaging the spectrumdisplay will eliminate theRandom Jitter (RJ) component.And since only one lane of data is being generated with all otherlanes “quiet,” the averagedspectrum display will not includeany correlated periodic jitter dueto coupling from adjacent lanes of traffic. You should see severalfrequency spurs indicating thepresence of ISI, DCD, andpossibly uncorrelated PJ. Storethis averaged spectrum displayfor later reference.

Now turn on all lanes of traffic in your system using the samepattern length for each lane oftraffic and perform the sameaveraged jitter spectrummeasurement as just described.Jitter contribution fromcorrelated periodic jitter will addto the height of the ISI & DCDspurs relative to the previousmeasurement that did not includecorrelated periodic jitter (secondtrace up from bottom). If there isa significant difference in spurheights from these twomeasurements, then this is anindication of the presence ofcorrelated periodic jitter.

To measure the peak-to-peakamplitude of correlated PJ, use the techniques previouslydiscussed to measure the amountof DJ with the system generatingjust a single lane of data traffic,

16

Analyzing a Spread Spectrum Clock

Jitter on a data signal is basicallyunwanted phase modulation.However, some serial datastandards actually specifyintentional modulation of theserial data with Spread SpectrumClocking (SSC). You can use jitter analysis to measure theshape and frequency of thisdesigned-in modulation.

Figure 18 shows an example ofmeasuring the spread spectrumclock when performing jitteranalysis on a 2.5 Gb/s datastream. By setting up the jitter

analysis to display the jitter trend waveform based on afixed-frequency clock, the resultsof the jitter measurement will be dominated by the spreadspectrum clock as shown by thepurple trace (bottom) in thisscreen-shot. We can easilymeasure the frequency of thisclock using either the scope’smarkers or automatic parametricmeasurements. In this case, wemeasured 32.9 kHz.

Just as memory is important fordetecting low-frequency PeriodicJitter (PJ), having sufficientmemory is also important forperforming spread spectrum

At this point you can measure the contribution of these jittercomponents by performing an ac RMS measurement on thiswaveform. Note: Random Jitter(RJ) is always measured in termsof an RMS value since it istheoretically unbounded. In thisexample we have measuredapproximately 5.7 ps rms ofrandom jitter. This amount ofjitter may not sound like much,but when converted into apeak-to-peak value based on a BER of 10-12, 5.7 ps rms RJtranslates into approximately80 ps of peak-to-peak jitter.

If you believe that theuncorrelated periodic jittercomponent is significant, you can use techniques previouslydiscussed to detect and measurethe uncorrelated periodic jittercomponents. But rather thanmeasuring the peak-to-peakamplitude of uncorrelated PJ component as previouslydescribed, measure the ac RMSvalue over a single period. Youcan then use the square root ofthe sum of the squares formula to separate out the uncorrelatedperiodic jitter component fromthe RMS measurement thatincludes both random components(RJ + uncorrelated PJ). As afirst-order approximation, theresults will be the RMS value of just the random jitter (RJ) component.

Figure 18. Measuring the Spread Spectrum Clock.

17

if you wanted to analyze theunwanted modulation (jitter) inthis 2.5 Gb/s data stream, youwould use a “golden” PLL-typeclock with a specific loopfrequency in order to filter-outthe intended low frequencymodulation of the spreadspectrum clock.

If your system is based on anembedded PLL clock, then tomeasure system jitter (not SSC) you should always use a PLL-type clock as your reference,regardless of whether yoursystem uses a spread spectrumclocking scheme or not. Thismeans that low-frequencyperiodic jitter (PJ) will also be filtered out of the jittermeasurements. But this is okay.The hardware PLL clock in yoursystem will also filter out any lowfrequency period jitter. Whenmaking jitter measurements witha software-generated clock, wewant our software-generatedreference clock to emulate the system’s embedded hardware clock.

Conclusions

Some real-time jitter analysispackages give answers in terms of the amount of total jitter thatmay be present in your system.This can be important fordetermining if your high-speeddigital system meets a particularworst-case jitter and eye-openingspecification. But knowing how much random jitter anddeterministic jitter is in yoursystem usually doesn’t give you a clue as to where it is comingfrom. The key to finding sourcesof jitter lies in the ability totime-correlate jitter measurementresults with high-speed serialdata signals, as well as otherpossible sources of uncorrelatedperiodic jitter. A real-timeoscilloscope with jitter analysisalong with the stimulus-responsetechniques described in this paper meet that criticaltime-correlation requirement torelate jitter trend measurementresults to measured signals. Onceyou are able to time-correlateparticular real-time timing errormeasurements to particular bitswithin a serial data pattern, it usually becomes a routinetroubleshooting task to solve yourdeterministic jitter problems.

clock measurements. Capturing a minimum of one period of a30 kHz spread spectrum clock at a sample rate of 20 GSa/srequires a minimum of 667 K of acquisition memory.

You should be cautioned that not all real-time jitter analysis iscapable of measuring the spreadspectrum clock. When timingerrors exceed one Unit Interval(UI), which is equivalent to oneperiod of the software-generatedreference clock, some jitteranalysis will “snap-back” andmeasure the timing error of thedata edge relative to the nearestclock edge. Measuring the SSC requires that time errormeasurements be able to exceedone UI. The jitter analysissoftware must be able to trackUIs in the serial data pattern andthen measure the timing errorrelative to the appropriate clockedge... not the nearest clock edge.

Although we used afixed-frequency softwaregenerated clock as a reference forthis particular SSC measurement,

18

54850 Series Scopes and 1130 Series Probes

The highest-performance end-to-endmeasurement system availableExperienced scope users knowthat their measurements are onlyas good as their probing system.And as bandwidth increases, it isincreasingly important to makesure you are measuring yourcircuit, not your scope probe.Nothing is more frustrating thanchasing down an apparent designproblem, only to find that it wascaused by an inferior scope probe.Together, the newest Infiniiumscopes and the breakthroughInfiniiMax high performanceprobing systems offer anend-to-end measurement solutionwith unmatched performance,accuracy and connectivity. Theresults are measurements you cantrust and better insight intocircuit behavior.

InfiniiMax high-performance activeprobe systemThe innovative InfiniiMax probingsystem provides either differentialor single-ended probing solutionsfor the most demandingmechanical requirements,without sacrificing performance.A flat frequency response overthe entire probe bandwidtheliminates the distortion andfrequency-dependent loadingeffects that are present in probesthat have an in-band resonance.

Performance-enabling technologyWhen you need to makemulti-channel measurements onprojects that use subnanosecondlogic, you need powerfulinstruments. These new Infiniiumoscilloscopes maintain their fullsampling performance of 20 GSa/son all channels, so you can makecritical timing measurements at the full performance of the oscilloscope.

Figure 19. Agilent’s 54855A 6 GHz Real-time Oscilloscope.

Features• 6 GHz bandwidth real-time

oscilloscopes with 20 GSa/ssample rate on all fourchannels simultaneously

• Up to 1 Mpts MegaZoom deepmemory at all sample ratesand 32 Mpts MegaZoom deepmemory at 2 GSa/s and slowersample rates

• Electronic attenuatorseliminate the reliability andrepeatability concernsassociated with mechanicalattenuator relays

• Trigger jitter as low as1.0 ps rms

• Easy-to-use, easy-to-understandjitter analysis option

• E2688A Serial DataAnalysis/Mask Testing withClock Recovery

• Serial ATA Signal QualityCompliance Test

• Each InfiniiMax probeamplifier supports bothdifferential and single-endedmeasurements for a morecost-effective solution

• Unrivaled InfiniiMax probingaccessories support browsing,solder-in, and socket usemodels at the maximumperformance available

• The 7 GHz InfiniiMax 1134Aprobe and E2668A/E2669Aconnectivity kits (all soldseparately) are recommendedfor the Infiniium 54855Aoscilloscope

• USB 2.0 Compliance TestOption

• Communications Mask Test Kit• VoiceControl Option

19

Related Literature

Publication Title Publication Type Publication Number

Jitter Solutions for Digital Circuits Brochure 5988-8427EN

Jitter Measurement Solutions CD 5988-9350EN

Measuring Jitter in Digital Systems Application Note 1448-1 5988-9109EN

Signal Integrity Solutions Brochure 5988-5405EN

Signal Integrity Solutions CD 5988-6915EN

81133A and 81134A 3.35 GHzPulse/Pattern Generators

The need for pulse and patterngeneration is fundamental todigital device characterizationtasks. The ability to emulate thepulse and pattern conditions towhich the device will be subjectedis essential. This emulationshould include both typical andworst case conditions. Accurateemulation requires superlativesignal integrity and timingperformance along with fullcontrol over parameters thatallow specific worst case testing.

Visit Our Websites

Signal Integrity ApplicationsCentral: www.agilent.com/find/si

Signal Integrity eSeminar Series: www.agilent.com/find/sigint

Jitter Measurement Solutions: www.agilent.com/find/jitter

Setting StandardsThe Agilent 81133A and 81134A3.35 GHz Pulse Generatorsprovide programmable pulseperiods from 15 MHz (66.6 ns) to 3.35 GHz (298.5 ps) pulsecapability on all channels. Atthese frequency ranges thetransition time performancebecomes critical; less than 60 psis specified and with a RMS jitterof 1.5 ps typical signal quality isassured. The Delay Control Inputand the Variable Crossover Pointfunctionalities allow emulation ofreal world signals by adding jitterto clock or data signals or bydistorting the ‘eye’ for eyediagram measurements.

Features• Frequency range from

15 MHz to 3.35 GHz• Low jitter• LVDS applications can

be addressed with outputlevels from 50 mV to 2.00 V

• PRBS from 25-1 to 231-1• Fast rise times (20%-80%)

< 60 ps• Delay Modulation

(jitter emulation)• Variable Cross-Over Point

(eye deformation)• 8 kBit data pattern memory,

RZ, NRZ, R1, Burst capability• 12 MBit extended pattern

memory • Graphical User Interface• All Inputs and Outputs are

SMA connectors

Figure 20. Agilent’s 81134A 3.35 Gb/s Pulse/Pattern Generator.

Agilent Technologies’ Test and Measurement Support, Services, and AssistanceAgilent Technologies aims to maximize the value you receive, while minimizing your risk andproblems. We strive to ensure that you get the test and measurement capabilities you paidfor and obtain the support you need. Our extensive support resources and services can helpyou choose the right Agilent products for your applications and apply them successfully.Every instrument and system we sell has a global warranty. Support is available for at leastfive years beyond the production life of the product. Two concepts underlie Agilent's overallsupport policy: "Our Promise" and "Your Advantage."

Our PromiseOur Promise means your Agilent test and measurement equipment will meet its advertisedperformance and functionality. When you are choosing new equipment, we will help youwith product information, including realistic performance specifications and practicalrecommendations from experienced test engineers. When you use Agilent equipment, wecan verify that it works properly, help with product operation, and provide basic measurementassistance for the use of specified capabilities, at no extra cost upon request. Many self-helptools are available.

Your AdvantageYour Advantage means that Agilent offers a wide range of additional expert test andmeasurement services, which you can purchase according to your unique technical andbusiness needs. Solve problems efficiently and gain a competitive edge by contracting withus for calibration, extra-cost upgrades, out-of-warranty repairs, and on-site education andtraining, as well as design, system integration, project management, and other professionalengineering services. Experienced Agilent engineers and technicians worldwide can helpyou maximize your productivity, optimize the return on investment of your Agilentinstruments and systems, and obtain dependable measurement accuracy for the life ofthose products.

By internet, phone, or fax, get assistance withall your test & measurement needs

Online assistance:www.agilent.com/find/assist

Phone or FaxUnited States:(tel) 800 829 4444

Canada:(tel) 877 894 4414(fax) 905 282 6495

China:(tel) 800 810 0189(fax) 800 820 2816

Europe:(tel) (31 20) 547 2323(fax) (31 20) 547 2390

Japan:(tel) (81) 426 56 7832(fax) (81) 426 56 7840

Korea:(tel) (82 2) 2004 5004(fax) (82 2) 2004 5115

Latin America:(tel) (305) 269 7500(fax) (305) 269 7599

Taiwan:(tel) 0800 047 866(fax) 0800 286 331

Other Asia Pacific Countries:(tel) (65) 6375 8100(fax) (65) 6836 0252Email: tm_asia@agilent.com

Product specifications and descriptions in thisdocument subject to change without notice.

© Agilent Technologies, Inc. 2004Printed in USA June 25, 2004

5988-9740EN

www.agilent.com

www.agilent.com/find/emailupdatesGet the latest information on the products and applications you select.

Agilent T&M Software and ConnectivityAgilent's Test and Measurement software and connectivity products, solutions anddeveloper network allows you to take time out of connecting your instruments to yourcomputer with tools based on PC standards, so you can focus on your tasks, not on yourconnections. Visit www.agilent.com/find/connectivity for more information.