[time-nuts] PICTIC II ready-made?

David davidwhess at gmail.com
Fri Apr 27 18:17:45 UTC 2012


On Fri, 27 Apr 2012 16:30:11 +0200, Attila Kinali <attila at kinali.ch>
wrote:

>On Wed, 25 Apr 2012 23:30:45 -0500
>David <davidwhess at gmail.com> wrote:
>
>> If you add a second lower current source or sink, then you can get
>> away with a LM311 class comparator and one fast timer channel in the
>> microcontroller.  The input pulse width charges the capacitor and the
>> timer counts how long it takes to slowly discharge.  Since the
>> conversion is integrating instead of sampling, it has better noise
>> immunity.
>
>Yes, a dual slope time strecher would work too. I'm not sure, but
>i would guess this aproach would be a lot more limited by the noise
>and device variations.

It would be a lot more immune to noise.  Both integrating and sampling
designs suffer from the same device variations which can be removed
through self calibration.

>Usually a timing input of an uC runs with a counter in the region
>of 100MHz max, ie +/-5ns resolution. To get to 50ps, one would need
>to stretch it by a factor of 100 at least, better 1000 to get some
>headroom for calibration in software. This means that the currents
>have to have a factor of 1000 in between. Using a charge current
>somewhere between 10 to 100mA would yield to a discharge current
>between 10 to 100uA. Keeping the two current sources stabile
>enough for the ratio to stay stable would be already quite an
>acheivment. Also keeping the leakage currents at bay would be
>quite some feat...

That is about the performance level of the Tektronix 2440 delay time
counter.  The counter only runs at 40 MHz but both edges of the 500
MHz sampling clock are used with two integrators so that metastability
can be detected and resolved.  The charge current is fixed at about
25mA and the discharge current is set during self calibration to
maintain a 1250:1 ratio at about 20uA.

Stability should not be a problem in the analog design when self
calibration is used and that is required at higher performance levels
anyway.  Even the high offset voltage and bias current of the bipolar
technology LM311 only contributes offset and gain error which is how
they got away with 100pf of integration capacitance.

>In contrast to that, a 16bit ADC is dirty cheap and a 24bits are readily
>available. I haven't had a look at it yet, but if the capacitive charge
>redistribution ADCs simplifiy the circuitry that much as Bruce has said, then
>you could get "easily" 16-18bit resolution. Combine that with a 100MHz
>reference clock, then you get a nominal resolution 150-40fs(!). 
>Acheiving 10ps resolution should be then a piece of cake and 1ps possible.
>(yes, i know that 10ps is not that easy...)

Charge redistribution ADCs by design have a built in sample and hold
which can simplify external circuitry and like delta-sigma converters,
they can be built on a digital logic process.  In this case, the
simplification is in comparison to non-sampling converters where the
signal level has to be constant during the conversion cycle for valid
results.

The advantage with the dual slope design is that it is integrating so
high frequency noise is ignored.  Controlling noise in a
microcontroller sampling ADC even at the 10 bit level is a significant
challenge.  In a conservative design, I usually start by figuring the
loss of one bit do to DNL and another bit do to noise.  If you want
better performance, the ADC either needs to be integrating or external
where noise can be better controlled.

I have been looking at a better than 10ps performance design but not
primarily for GPS timing applications.  I am more interested in
equivalent time sampling and high bandwidth sequential or random time
sampling.  The later can not use an integrating converter because of
sampling rate requirements.



More information about the time-nuts mailing list