[time-nuts] Experiment in lowering the TAPR TICC noise floor

[Attila Kinali]

On Sat, 03 Oct 2020 10:37:59 +0200
Matthias Welwarsky <time-nuts at welwarsky.de> wrote:

> When I started to look more into the software side of the TICC and especially 
> the ominous "time dilation" parameter, I set up an experiment where I feed the 
> same event into both channels of the TICC, for evaluating the sensitivity of 
> the measurements to this parameter (spoiler: there is a measurable influence 
> but it's not as critical as I originally thought).

That is to be expected. There are two resons for this:

First, the major limit to the measurement is the noise within
the TDC7200. If you want to get lower, then you have to reduce
this noise. If you look at Figure 17 in the TDC7200 manual, you
will see that the noise of the TDC is highly dependent on the
length of the measurement. Shortening the measurement will
decrease the noise. For this you need to use a higher clock
of the stop signal to measure against, than the 1ms that the TICC
does. But that will not work with the Arduino. You can get around
this if you use a faster µC like an STM32F4. See Tobias Pluess GPSDO
design for an example how to do this.

Second, both inputs of the TICC measure against the same divided
1kHz clock with a modified half-Nutt interpolator. I.e. most of
the measurement time will be common to both input signals and thus
most of the noise seen due to the TDC and the reference clock are
common.


On Wed, 07 Oct 2020 18:34:00 +0200
Matthias Welwarsky <time-nuts at welwarsky.de> wrote:

> the noise is likely not white, but it really depends on what is the dominant 
> noise source in the system. I guess there is some correlation but still enough 
> entropy to make a difference. I'll try with different cable lengths next to 
> see if it makes a measureable difference, but ideally you'd use two TICCs and 
> two non-coherent reference clocks. But they'd need to be somehow frequency 
> locked.. You'd need some mechanism that causes enough jitter to break the 
> correlation. A delay line controlled by some noise source?

Adding noise will not break any correlation. It will only mask it.
I.e., the correlation will pop up once again, when you start
using methods to remove the added noise.

Adding noise helps only if your noise is mostly quantization noise, 
then it acts as a dithering mechanism which allows you to average 
over the quantisation (and added) noise, which wouldn't be possible
otherwise.


			Attila Kinali
-- 
<JaberWorky>	The bad part of Zurich is where the degenerates
                throw DARK chocolate at you.