[Gb-ccb] Minutes from 07jul03 videocon

[Martin Shepherd]

Note that in addition to comments regarding the minutes of the
teleconference, there are some important points in this email,
including a decision on what type of low-pass filter to use.

On Wed, 9 Jul 2003, Brian Mason wrote:
> Background: We held the videocon in order to provide feedback on a
> number of specific issues that Caltech had raised now that they are
> getting on with the hardware design.  The main issues were: the
> overall architecture change (to a faster, lower-resolution ADC further
> up in the signal chain, followed by low pass filters and integrating

Actually the ADCs are preceded by low-pass filters, not followed by
them.

>...
> Clarifying our earlier definition of "fast" we said the step response
> should be 0.5 usec, settling to -84dB at 1 usec (-84 dB is a part in
> 4000 that is, 12 bits).

Actually, 12 bits is 72dB, not 84dB, but the 1 in 4000 is essentially
correct (ie. 2^12=4096).

>....
> backends which I didn't quite get down clearly (the signal is down by
> 15 dB at the nyquist frequency, which results in aliased signal being
> 30 dB down?)

Curious. How do you get from 15dB aliased signal strength to 30dB
down? I would have thought that it would simply be 15dB down. If I am
wrong, please tell me, since a doubling of the filter dB stop-band
attenuation would be great :-).

>-- and wonders whether the aliased stuff is signal or
> noise (worst case: some sort of systematic noise which doesn't average
> down).

Not systematic noise, simply higher-frequency receiver and detector
noise.  The key to understanding this is to realize that the time
window (tw) during which the ADC samples its input signal is shorter
than the sampling period (T). Without an anti-aliasing filter, the
duration of this sampling window sets the averaging time, whereas with
an anti-aliasing filter the time-constant (tf) of the filter sets the
averaging time.

A perfect anti-aliasing filter thus improves the SNR seen at the input
of the ADC by the factor:

  sqrt(tf/tw)

However a low-pass filter with poor stop-band attenuation is like a
filter with a smaller time-constant than that implied by the cutoff
frequency, and thus the SNR is degraded.

I have been researching filter choice for the last couple of days, and
have reached the following conclusions.

When choosing a filter for the CCB, it is important to keep in mind
that the primary goal of this filter is not anti-aliasing, but
averaging the input signal for a strictly bounded time period. In
other words the ideal filter for this application would have the
time-domain impulse response of a top-hat function, which would
dictate a sync function filter. Such a hypothetical filter (which
doesn't exist) would be a disaster in the frequency domain, since it
would ring very badly, well beyond the Nyquist frequency. As such one
needs a compromise filtering function. The best readily available
option is the Bessel filter, which is designed to exhibit virtually no
overshoot or ringing in the time-domain, and has a fast step-response
time, due to it having approximately constant delay as a function of
frequency. To counter the relatively poor frequency response of this
type of filter, I suggest that instead of using a 5Mhz anti-aliasing
filter, we use a 2MHz low-pass Bessel filter of 8 or 10 poles. At the
Nyquist frequency of 5MHz both 8 and 10 pole Bessel filters have
attenuations of about 20dB. At the sampling frequency of 10MHz, the
attenuation of a 10-pole Bessel filter is about 70dB.

Since there would be nothing detectable by the ADC beyond 10Mhz using
one of these filters, we know that the increase in the SNR due to its
imperfect frequency response is less than 40% compared to the ideal
SNR. In practice, given that the filter is already down by 20dB at the
Nyquist frequency, the noise degradation will be a lot less
significant than this.

Unlike most other types of filter, the theoretical time needed by a
2MHz Bessel filter to settle to virtually any practical accuracy is
slightly over the reciprocal of its cutoff frequency. Thus a 2MHz
low-pass Bessel filter should settle within just a little over 0.5us.

So, the specification for the filter that I would like is a Bessel
filter with at least 8 poles, and a cutoff frequency of 2Mhz. Suitable
filters can be bought from TTE, at:

 http://www.tte.com/

The base of a 2MHz PCB-mounted 8-pole Bessel filter from TTE has
dimensions of 2.4" by 1". Thus 16 of these filters would occupy
upwards of 38 square inches of the PCB (ie. 6"x6"). The height of the
filter is 0.5".

> The full-scale voltage of the ADC being considered is 2.5 Volts.

This needs clarification. When operated differentially, each of the
two differential inputs of the ADC can swing over a 2.5V range between
the limits 1.25V and 3.75V (ie the common-mode voltage is 2.5V). The
optimum ADC noise performance is achieved for signals at the
common-mode voltage.  Ideally the signal would be offset to place the
signal level that is generated by the receiver when there is no sky
signal, at the common-mode voltage.  In principle I could base such an
offset on the output of a trimmer-pot attached to the ADC reference. I
am still pondering whether to do this, or simply to place the 0V of
the detectors at the common-mode voltage.

> We
> indicated to Caltech that from our point of view, the signal levels
> are open to specification by them. The ADC is clocked at 10 MHz
> (5xNyquist).

Hum, the Nyquist frequency is the ADC sampling rate divided by
2. Maybe you meant to say 5 times the cutoff frequency of the low-pass
filter?

> Discussion of the control signals was somewhat less conclusive.  Will
> we use LVDS or TTL?

I thought that we had finally decided on opto-isolated TTL?

>...
> receiver.  Are both TTL and LVDS compatible with using the opto
> isolators?

I don't know of any opto-isolators that are designed to generate LVDS
outputs.

> Not clear.  How fast can the opto-isolators be driven?
> Martin will look into the last question.

As reported yesterday via email, suitable devices with 75ns
propagation delays are readily available (and cheap).

> ...
> are combined and sent.  A sketch of the control signals would be
> helpful and GB will produce this.

My previous documents already contain these diagrams.

> As to interconnection: the general feeling was that more connectors
> rather than fewer was good: this permits empirically optimizing the
> grounding and bundling scheme, and presents more options for
> troubleshooting should they be needed.

That wasn't the impression that I got. My recollection was that we
settled on using 4 cables containing 4 pairs each, with the caveat
that if subsequently found to be necessary, we could always retrofit
the CCB and the receiver to use 16 individual cables.

> Martin asked if we have some video filters lying around which we could
> have a quick look at.  We do, and we might be able to have a look at
> them; failing that we could probably give some pointers.

I don't think that this will be necessary now. I am comfortable that
the Bessel filters that I described above will do the job.

Thanks.

Martin Shepherd  (mcs at astro.caltech.edu)