The 16 bit ADCs (at least the LTC lower sample rate ones ) tend to be a few dBc/Hz quieter.

However the difference is < 12 dBc/Hz.

Bruce

On Fri, 27 Oct 2017 09:35:08 +0800
"Li Ang" 379998@qq.com wrote:

Indeed a nice post, as usual for Andrew.

Which is close to the theoretical limit.

But be aware, that measurements close to the limit of thermal noise
will make your measurement go sour. There the noise of your splitter
will cause an anti-correlation effect and the measured noise will
suddenly drop way below thermal noise. Craig Nelson and Archita Hati
from NIST, Enrico Rubiola from FEMTO, Magnus from time-nuts and several
others have been discussing this for a couple of years now at PTTI,
IFCS and EFTF.
(e.g. https://www.nist.gov/publications/cross-spectral-collapse-anti-correlated-thermal-noise-power-splitters )

Beside the price? :-)
There are two things that limit the measurement here. One is the noise
of the ADC itself and the other is the apperture jitter. The former
is almost the same for both 14 and 16bit high speed ADCs. Unless you
go well below 30Msps, you will not gain much in noise performance
from going to 16bit.

The aperture jitter is more or less unrelated to sampling rate and
bit width but faster ADCs usually have lower aperture jitter,
as the designers take more care (because it becomes more important
for "normal" applications). This means, that if you are limited by
aperture jitter (which is the case for close in measurements below
1-100Hz probably), then the additional bits of a 16bit ADC will not
help you at all.

That said, it is probably worth trying what actually happens when
using a 16bit ADC instead of 14bit. If there are any students here
looking for a bachelor or master thesis project doing noise measurement,
feel free to contact me :-)

		Attila Kinali

--
You know, the very powerful and the very stupid have one thing in common.
They don't alters their views to fit the facts, they alter the facts to
fit the views, which can be uncomfortable if you happen to be one of the
facts that needs altering.  -- The Doctor

On 11/12/2017 03:57 PM, Attila Kinali wrote:

Quite respectable. Averaging may help some but...

This is a hard problem. Turns out that beyond the anti-correlation noise
from the splitter, isolation breaks in many interesting ways, creating
cross-talk which can or can not cause disruptions. My approach to this
is to steer the effects of the cross-talk away from the measurement.
This is however a hard balance to maintain in practice, but hopefully
will inspire new approaches.

Well, that is not completely true. As you average cross-correlations you
sense a thad more of the actual noise, so there is a small gain there.
Having too few bits will acts as a discriminator for how low levels
become visible, even if covered in noise. We've seen people make this
mistake before. Just because it's hidden in noise does not prohibit it
from being detected in a correlation process, and cross-correlation is
one such process.

I saw one such spectrum this summer that had exactly this fault, there
is a limit to how deep a cheap SDR goes, and the mistake was quickly
pointed out by someone with the bitter experience.

Noise tend to smooth over quantization steps, and hum, and I think you
recall my poor excuse of a poster-presentation on that topic, but to
make benefit it takes lots of samples.

So, anyway, there is some benefits in that extra bits. To make good use
of them one need to be a bit careful in processing.

Just do it. I'm sure I can contribute some to that if it would happen.

Cheers,
Magnus

On Sun, 12 Nov 2017 16:34:35 +0100
Magnus Danielson magnus@rubidium.dyndns.org wrote:

I would like to. But I don't have the time to do it myself.
I already have way too many projects running in parallel as it is.

		Attila Kinali

--
You know, the very powerful and the very stupid have one thing in common.
They don't alters their views to fit the facts, they alter the facts to
fit the views, which can be uncomfortable if you happen to be one of the
facts that needs altering.  -- The Doctor

Cross-spectral collapse due to thermal noise is basically Mother Nature complaining that we're trying to do something that's physically ill-advised, by trying to measure the noise in a system at levels near the thermal floor of the termination resistance that's necessary to define such a system in the first place.  Unfortunately the issue can affect measurements at levels several dB higher than the thermal floor, making it a genuine occupational hazard.

On the bright side, splitter topologies that use a single resistive load with no differential-mode termination have little or no vulnerability to this phenomenon.  Of course the common-mode termination imposes its own noise floor, but that's life in this particular universe.

That being said, it's still a bit of an awkward situation for designers of future test sets.  We can't use Wilkinson splitters anymore -- or at least we shouldn't, given what's been learned over the past couple of years.  As a result, some customers' best devices may suddenly start looking noisier than they did when measured on earlier instruments, or (worse) when measured on competing ones.  Those customers may not respond agreeably when advised to dunk the termination in liquid helium...

None to speak of.  The original TSC 5120A used 14-bit parts while the later test sets have used 16-bit parts, and the performance limits are essentially identical.  Once you go below a certain point, generally well below -170 dBc/Hz, you tend to run into a correlation floor that arises from one or more different causes unrelated to the quantization precision itself.  These effects seem to arise in and near the S/H stage and aren't improved by going beyond 14-bit precision.

-- john, KE5FX
Miles Design LLC