Jitter measured relative to the reference input (trigger on reference
input, measure synthesizer output). 2.5GHz conventional PLL, ring
oscillator, standard 0.18 digital CMOS process, ~75MHz reference.

-- Mike --

On Thu, 28 Feb 2008 07:59:41 -0800, Mike wrote:

Jitter measured relative to the reference input (trigger on reference
input, measure synthesizer output). 2.5GHz conventional PLL, ring
oscillator, standard 0.18 digital CMOS process, ~75MHz reference.

-- Mike --

What model scope is that?

Take a look at the flat parts of the waveform, and see how fuzzy they
are. Your scope is running at 10 mV/cm, and if there's even millivolts
of vertical noise from the scope or ground loops or something,
vertical noise will look like time jitter.

Your jitter may well be less than the scope reports!

John

John Larkin wrote:
On Thu, 28 Feb 2008 07:59:41 -0800, Mike wrote:

Jitter measured relative to the reference input (trigger on reference
input, measure synthesizer output). 2.5GHz conventional PLL, ring
oscillator, standard 0.18 digital CMOS process, ~75MHz reference.

-- Mike --

What model scope is that?

Take a look at the flat parts of the waveform, and see how fuzzy they
are. Your scope is running at 10 mV/cm, and if there's even millivolts
of vertical noise from the scope or ground loops or something,
vertical noise will look like time jitter.

Your jitter may well be less than the scope reports!

I only know the scope as a Tek Communications Signal Analyzer (I suspect
there's a particular model number, but I don't know what it is offhand).
The signal amplitude is around 1Vpp; I zoomed in on the rising edge to
make the measurement.

As I recall, there are two display modes for the analyzer - points or
continuous segments. We generally run in the points mode, which hides
any visible amplitude noise. The number of hits, just over 500k, is the
number of points that fall within the thin light-blue rectangle along
the center line.

In addition to amplitude variations, the delay through the circuit
varies with supply and temperature. Our supply is pretty stable, but the
temperature of the lab varies over time - by a few degrees during the
day and much more when the HVAC shuts off at night. Those delay
variations cause the overall delay to change, shifting the output
relative to the input. When capturing large numbers of points, it wasn't
uncommon for the A/C to cycle on or off. The result would be a very
smeared trace, and useless statistics. At 4ps per division, you could
see the effects of temperature by simply fanning the part.

The reason I wanted to take so many points is that there continues to be
a large number of engineers who believe that the Pk-Pk jitter number is
constant, not a function of the sample size. So, I was really after
Pk-Pk jitter for sample sizes from 1k points to 10 million. I never got
past 500k - we couldn't keep the system stable long enough to make the
measurement. From 1k to 500k, though, a plot of log(N) vs Jpk-pk^2
produced a nicely linear plot (more linear for larger N), just like
theory predicts. Yes, for the nitpickers that read this, I'm assuming a
Gaussian distribution.

-- Mike --

On Fri, 29 Feb 2008 08:46:59 -0800, Mike wrote:

John Larkin wrote:
On Thu, 28 Feb 2008 07:59:41 -0800, Mike wrote:

Jitter measured relative to the reference input (trigger on reference
input, measure synthesizer output). 2.5GHz conventional PLL, ring
oscillator, standard 0.18 digital CMOS process, ~75MHz reference.

-- Mike --

What model scope is that?

Take a look at the flat parts of the waveform, and see how fuzzy they
are. Your scope is running at 10 mV/cm, and if there's even millivolts
of vertical noise from the scope or ground loops or something,
vertical noise will look like time jitter.

Your jitter may well be less than the scope reports!

I only know the scope as a Tek Communications Signal Analyzer (I suspect
there's a particular model number, but I don't know what it is offhand).
The signal amplitude is around 1Vpp; I zoomed in on the rising edge to
make the measurement.

As I recall, there are two display modes for the analyzer - points or
continuous segments. We generally run in the points mode, which hides
any visible amplitude noise. The number of hits, just over 500k, is the
number of points that fall within the thin light-blue rectangle along
the center line.

In addition to amplitude variations, the delay through the circuit
varies with supply and temperature. Our supply is pretty stable, but the
temperature of the lab varies over time - by a few degrees during the
day and much more when the HVAC shuts off at night. Those delay
variations cause the overall delay to change, shifting the output
relative to the input. When capturing large numbers of points, it wasn't
uncommon for the A/C to cycle on or off. The result would be a very
smeared trace, and useless statistics. At 4ps per division, you could
see the effects of temperature by simply fanning the part.

The reason I wanted to take so many points is that there continues to be
a large number of engineers who believe that the Pk-Pk jitter number is
constant, not a function of the sample size. So, I was really after
Pk-Pk jitter for sample sizes from 1k points to 10 million. I never got
past 500k - we couldn't keep the system stable long enough to make the
measurement. From 1k to 500k, though, a plot of log(N) vs Jpk-pk^2
produced a nicely linear plot (more linear for larger N), just like
theory predicts. Yes, for the nitpickers that read this, I'm assuming a
Gaussian distribution.


Yup, CMOS has a bad positive TC of delay versus temperature, in the
vague ballpark of 1% of prop delay per degree C. Some of that can be
compensated out with a temperature sensor driving something that
affects delay, Vcc maybe, but that only gets you so far, 5:1 if you're
lucky.

The telecom boys call "jitter" to be the standard deviation of delay
measured over 0.1 seconds, and anything slower is "wander."

Sometimes a jitter contributor is deterministic, like power supply
ripple or crosstalk or something. That tends to make the pdf less
gaussian, so you'd have fewer extreme outliers.


John