LeCroy DDA 7 Zi series, Operator'S Manual

Page: 341 / 544

Operator's Manual
WP700Zi-OM-E-RevA
340
each edge of the extracted clock signal has four samples between the top and base. These samples are placed
proportional to phase, so that the edge crosses 0 at exactly 0 and 180 degrees VCO phase.
The phase steering target for the VCO is, roughly speaking, that the data transitions should happen on the falling
edge of the VCO output. To be more precise, we steer such that the VCO phase will be 180 degrees at the
sample where a zero crossing in the data is detected. Because the software VCO works on a sample-by-sample
basis, there is, on average, a half sample delay from the VCO falling edge's zero cross to the data zero cross. At
20 samples per T, this half sample error is 2.5% of T, not noticeable without zooming in. At 10 samples per T, it is
5% of T. The following small figure shows part of a zoom on a rising data edge and a falling clock edge sampled
at 500 MS/s (2 ns per sample). The horizontal scale is 2 ns per division. The samples are bold. Note that the data
crosses zero at 1 ns after the falling edge of the VCO output crosses zero. This is the expected result.
The signal used is a 4.36 MHz square wave, which has a
transition every 3T when 1/T = 26.16 MHz. During the first 50 μs
or so, the phase settles in from initial startup. After that all the
zero crossings are half a sample apart, as shown in this picture.
JitterTrack of Frequency (requires JTA option) of the extracted
clock. The startup frequency was correct to within a few kHz,
and the PLL did not slip. It is possible that the starting frequency
was precise but the starting phase was not; the effect would be
the same. JitterTrack shows frequency as a function of time.
The vertical scale is 20 kHz per division; the cursor is positioned
at 27.107 MHz. The horizontal scale is 0.1 ms per division.
How the Starting VCO Frequency & Phase Are
Determined
The PLL's VCO is started at a frequency of 1/T. Due to the accuracy required, T is determined in two steps. The
first step produces an estimate of T starting with very few assumptions. The second step starts with the estimate
of T and refines it. The information used in both steps is the sample at which a transition (through zero) occurred
in the sliced data, for up to the first 2000 edges. If the source waveform has less than 2000 edges, the accuracy
of this procedure may be reduced.
If the source waveform has less than 50 edges, the instrument does not even attempt to estimate T. The PLL will
start at 3.19024 * LP fc. Because of the low bandwidth of the PLL, it does not make much sense to try to extract
the clock from a very short waveform; the PLL will not have time to react.
The first step calculates the width of the first (up to) 2000 pulses, sorts the widths, and finds the first three peaks
in the distribution of widths. The distribution is "smoothed" by a five-bin wide boxcar filter to prevent small local
events from misleading the peak detection. This is the primary reason why the signal must be over-sampled by
greater than 10x. The distribution of widths is similar to a histogram of pwid (pit width) on "leveled" output of the
ODATA function, using a threshold of 0.0 mV and measuring All widths. The spacing between the peaks is
approximately T, close enough to determine the lowest nT. The instrument calculates the estimate of T from the
means of the first three peaks, which are assumed to be lowest n, lowest n + 1, and lowest n + 2 (i.e., 3T, 4T, and
5T). This estimate is generally good to better than 1%.
The second step uses the location of the first (up to 2000) transitions, in order. It uses the estimate of T to
calculate n between each pair of same-polarity edges. If the estimate is within 1%, we have at least 50% margin.
A 50% margin occurs if a pair of same-polarity edges is 25T apart. On a good waveform, the count is likely to be