28 February 2011
Design margins continue to tighten as data rates increase and unit intervals (UI) shrink. However, PC board materials remain extremely ‘lossy’ in order to cut costs.
Consider that for signals running at 10 Gbit/s, the unit interval is now only 100 ps before any ‘real world’ effects have been added. This is to say that the eye has been shrunk to less than 100 ps.
The jitter or noise of the design must be cut in half, yet in most cases, the same board material (for example, FR4) is used from previous designs.
Therefore, all the jitter and noise improvements must be done in design.
Test and measurement equipment, such as real time oscilloscopes, only further exasperate this problem as they add noise and jitter that is not part of the design.
When equalisation and de-embedding techniques are used, the jitter and noise added via test and measurement equipment is made even worse. While these techniques help to avoid switching board material and can open even the most tightly closed eyes, they do not reduce oscilloscope intrusion. Instead, they actually increase their effect.
Real time oscilloscopes have important limitations including noise floor, jitter measurement floor, and bandwidth. All of these limitations erode crucial margins in design.
No longer can oscilloscopes and their signal integrity specifications be ignored as insignificant contributors to design margins.
Even after purchasing an oscilloscope with the ‘best’ signal integrity specifications, knowledge of oscilloscope hardware and software can be used to further maximise design margins.
There are a number of key specifications to consider when evaluating oscilloscopes.
Maximising design margins begins with the initial evaluation of the oscilloscope to be purchased. Every specification can, and will, impact margins.
In addition to the hardware specifications, knowing the trade-offs relevant to the different oscilloscope vendors’ tools such as jitter, equalisation, and waveform transformation tools will affect margins.
Additionally, probing is extremely important too, so knowing the hardware performance of the probes can be the difference between passing and failing.
Combining these three key areas of evaluation and choosing the oscilloscope that best meets a company’s needs will ultimately save time and money.
Literature from oscilloscope vendors will often differ on which specifications are the most important for evaluation. These specifications include noise floor, intrinsic jitter, bandwidth, effective number of bits, and sine wave repeatability.
All of these specifications have merit when maximising margins, but for margin testing of a digital signal’s real time eye, the oscilloscope’s noise floor is one of the most important specifications.
Oscilloscope noise erodes eye height and impacts eye width as it will erode rise times. Vendors will specify noise floor in two different ways. Firstly, vendors set the oscilloscope to 50 mV/div and disconnect all channels. Then they turn on a vertical histogram of the noise and take the measurement. The measurement is compared against other oscilloscope vendors with the same bandwidth settings.
For example, at 25 GHz the Agilent Infiniium DSAX92504A measures the industry’s lowest noise floor of less than 1.8 mV assuming that there is no offset and the signal is positioned in the centre of the screen (figure 2).
This compares to a noise floor of 2.8 mV at 20 GHz for other oscilloscope vendors. Essentially, the DSA92504A yields 25 GHz with 45% less noise than any comparable 20 GHz oscilloscope on the market.
The second method used to measure noise floor is to take into account that some vendors use eight divisions on screen whilst others use ten.
This type of measurement is called Noise Percentage on Screen, and is taken the same way as the first method, except that the noise number is divided by the total mV possible on the screen.
For instance, with eight divisions at 50 mV/div, the DSAX92504A would have a Noise Percentage on Screen of 2 mV divided by 400 mV, or an industry best of 0.5% at 25 GHz.
In addition to noise with no offset, measuring the noise with offset added is important. To take this measurement, move the position of channel one’s signal (in this case just noise) to one division higher than centre on screen.
The histogram now reads the new noise measurement, which in the case of the DSAX92504A is 2.10 mV, or 0.52% for the percentage on screen.
Some oscilloscopes will vary by as much as 40% with this measurement change. In fact, the leading 20 GHz oscilloscope measures 3.3 mV at these conditions. Variation this great could mean that a signal with no offset passes a mask test, while a signal with offset fails due to the oscilloscope. Failing because of the noise variation of the oscilloscope is unacceptable.
Of course, there is more to an oscilloscope’s accuracy than just its noise floor; oscilloscopes have an inherent jitter. As UIs continue to shrink, oscilloscope jitter has an even greater influence on jitter measurements.
Even more important is that the intrinsic jitter measurement floor of the oscilloscope translates into random jitter of the device. It is crucial to not just take the oscilloscope’s influenced jitter specification listed in the data sheet. Some oscilloscope vendors will specify only the absolute lowest jitter number; assuming essentially no noise floor influence from the oscilloscope.
While this jitter (known as sample clock jitter) is important, it does not reflect the reality of what actual testing will reveal. The jitter measurement floor (sample clock jitter combined with noise influence due to slew rate) is influenced by the oscilloscope noise floor.
The slower the rise time (slew rate) being measured, the more influence the noise floor has. It is important to see the oscilloscope’s jitter measurement floor under many conditions rather than accepting the vendor’s data sheet.
Consider an oscilloscope with a jitter measurement noise floor of 180 fs (such as the DSAX92504A) and a noise floor of less than 2 mV. Then compare this to an oscilloscope with over 400 fs of jitter and a noise floor of greater than 3.5 mV.
If the slew rate is slow, for example 100 ps (10/90%), the oscilloscope’s noise will have a major influence on the final DUT results.
The measurement difference resulting from the difference in the jitter noise floors listed above could correspond to more than one picosecond of random jitter. This additional picosecond is then multiplied by fourteen for the total jitter measurement.
Ultimately, this means the oscilloscope is adding 14 ps of total jitter to the device under test. This measurement difference could mean that one oscilloscope passes while the other one fails the device under test.
Measuring the jitter measurement floor of an oscilloscope can be a fairly simple process depending upon the level of sophistication needed. By using a high accuracy sine wave source, the sine wave can be measured through the oscilloscope.
For the greatest accuracy, the results of the real time oscilloscope should be compared with the DCA-j (which has less than 50 fs of jitter measurement floor). When the sine wave is displayed, the time interval error measurement of the oscilloscope should be turned on, as shown in figure 2.
By specifying the constant clock recovery as a sine wave, it is easy for the clock recovery to lock on to the given data rate.
The final important component of measurement accuracy is the bandwidth of the oscilloscope. To accurately measure an edge, there must be enough bandwidth to give a truthful representation of the edge.
The edge will appear much slower than it actually is if an oscilloscope does not have enough bandwidth. The slower representation of the edge will result in higher jitter than actually exists in reality.
Dr. Howard W. Johnson has written a book on this topic entitled High-speed Digital Design – A Handbook of Black Magic, in which he refers to a key frequency component as the ‘knee’ frequency (fknee).
All fast edges have an infinite spectrum of frequency components, but there is an inflection (or ‘knee’) in the frequency spectrum of fast edges, where frequency components higher than fknee are insignificant in determining the shape of the signal.
When a value for fknee is determined, it is important to understand the frequency response of the oscilloscope. If an oscilloscope has a response that rolls off very quickly (a so-called ‘brickwall response’), it can accurately represent faster edges with less bandwidth. However, if an oscilloscope has a more Gaussian response, it will require more bandwidth.
If the device has a rise time of 25 ps (10/90%), then fknee is equal to 20 GHz.
Plugging this knee frequency into the equations above shows that for 10% accuracy, 24 GHz of oscilloscope bandwidth is required (assuming the oscilloscope exhibited a true ‘brickwall response’). In reality, oscilloscopes never achieve a true brickwall response and tend to have a response inbetween Gaussian and Brickwall, as shown in figure 5.
Contact Details and Archive...