Which jitter measurement is correct?
As you might have read in the last entry of 2011, Jit is sharing the blogger role in 2012 and is giving a number of us what we asked for – the chance to blog about signal integrity challenges. Consider us scope gurus in waiting.
My name is Randy White and I get to kick off the new year. I’m the serial applications technical marketing manager at Tektronix. This is such a great role because I not only get to work with the engineers building advanced instruments, but I also interact with designers like you who are putting the tools into practice. It provides a unique perspective. I also travel around and deliver seminars on high-speed serial measurements and am actively involved in working groups for high-speed serial standards. I hold a BSEE from Oregon State University in Corvallis.
Recently a question came up that I thought would be of interest to those of you working with high speed serial designs. Of the many jitter results, which one is correct? Why do I get different results from run to run or between different instrument vendors? Mark Guenther, a colleague of mine and an expert in jitter separation algorithms, has the answers to these questions.
Over the next few weeks I’ll be sharing his insights into how various jitter models differ and how they can relate to one another. To provide more background the original question was, “When using a BER= 1×10-12 level why don’t my RJ and DJ results follow the generally accepted formula of TJ = Dj + 14xRj?” This week, Mark sets up his explanation into why jitter results are different with a look at how actual jitter and Dual-Dirac jitter model relate to each other. Take it away Mark:
Does Tj@BER come from the Rj/Dj spectrum or from the Dual-Dirac model?
The answer is “from both.” The Dual-Dirac model is a simplified jitter model with only two parameters, the sigma of the RJ and the peak-to-peak span of the DJ. It begins with a measured bathtub curve, and the two parameters are adjusted in a particular way until a best fit is achieved between the Dual-Dirac bathtub and the measured bathtub. In the following graphic, the blue curve is a measured bathtub curve and the red curve is the Dual-Dirac curve that has been fitted:
This curve happens to use a Q scale for the vertical axis but you could also use either a log or linear scale. You can see that the blue (measured) curve has a lot of bumps and wiggles in the higher parts, corresponding to features of the deterministic jitter (DDJ, PJ). In contrast, the red curve is forced to have simple features by the simplicity of its model.
But where did the measured bathtub come from?
If you have a BERT, you can get the bathtub by sweeping the BERT’s sampling point across the unit interval and actually measuring the entire curve. If you have a real-time or sampling scope and a jitter-analysis package, you can get the curve by breaking the overall jitter into components and then reassembling those components in a way that allows extrapolation of the bathtub to low probability levels. In a jitter analysis tool, assuming you are using the repeating-pattern method, spectral analysis is used to get the blue curve (and to get the individual jitter components like PJ, DCD, etc.). Then the resulting bathtub curve, along with the above-mentioned curve-fit, is used to get the Dual-Dirac model parameters.
The following illustration gives another view of this process, and offers further insight into how the actual jitter and Dual-Dirac jitter relate.
As always, it would be great to hear from you about whether this type of content is useful to you, or if you have additional insights. Please add your comments below or drop me a note at scope-guru@tektronix.com.
Randy White, Serial Applications Technical Marketing Manager, Tektronix
Randy White is the Serial Applications Technical Marketing Manager at Tektronix. Randy has worked with various aspects of test and measurement solutions at Tektronix over the past few years. He has given seminars on high-speed serial measurements and is actively involved in many working groups for high-speed serial standards. He holds a BSEE from Oregon State University in Corvallis.
raymart commented:
Thanks Jon and Daniel. This is just what I needed.I am still very much a rdttiaional' illustrator and I am not sure I want to change that but I think it is important to try different things when possible. I will probably be one of those that does the drawing rdttiaionally and then uses photoshop for retouch or enhancing. The information you provided is easy to understand and follow. Thanks a million for putting that together.Now .how about a little lesson on channels and how to isolate complex images or better yet, what is the best method to use to isolate complex images. Thanks again!
Tom Gaudette commented:
I have also looked at this question and outside of the instruments you should look at the IEEE standard. The particular standard in play for what is jitter comes from the definition in the IEEE SA-181-2011. I went through this standard a couple years back (prior to the 2011 update) and found it quite interesting. From this standard you can define how you find the times of the crossing in four different ways with the last being user defined. Then once you find the crossing you can find the deviation. Depending on how short your signal really is each method will give you different results and from there different jitter measurements. But yet all of them are still accurate to the standard and therefore correct Jitter measurements. So if you are asking why do I get different results you really will need to understand what method each instrument is using, with them all still being within the standard.
Joe S commented:
Not addressing the different instrument models part of the question (although I could), I want to address the reason jitter results differ between runs: it's because Random Jitter is, by definition, RANDOM. Especially at the extremes of distribution, it can differ somewhat between sweeps. And, if any of the jitter is due to cross talk even though that's not random if the aggressor has different data it will have a different effect on different sweeps. One way to reduce these uncertainties is by taking a relatively long acquisition: you get better statistics from a million UI than from 10,000 UI. If your test equipment has a very high quality clock, that doesn't add its own jitter to long acquisitions, long acqusitions are unarguably a good thing. Thinking that every (even relatively short) acquisition should produce the exact same jitter results is simply incorrect thinking.
steve commented:
I agree, though at the sme time I am not sure if this has a simple answer. Different technologies have different sensitivities to jitter and specifically to the frequency range of sensitivity.
Since we are generally working with clocks and ADC's at our company, we use clock phase noise measurements using the Tektronix RSA5106A
Martin Rowe commented:
Every scope maker uses a different algorithm to calculate jitter. Thus, there are differences.
TifosiHendrix commented:
Am I missing something here or does this article not answer the original questions - why are jitter results different between instrument vendors and runs? I think there needs to be discussion about curve fitting techniques vs indepedent sigma techniques used by different vendors and the pros and cons. Also discussion on noise floor of instrumentation and how that contradicts with fundamental assumptions of the Dirac Delta model, along with the assumption that jitter is a stationary phenomenon (continuous data stream analysis not a reality) that leads to errors in measurement, is missing. Timing vs amplitude noise due to crosstalk and noise floor and inability to distinguish between that is another factor. Thanks.
