>
> True about not adding more usable data. But in this application what we're
> looking for is accurate measurement of the time when the analog waveform
> satisfies a certain condition. The more frequently we sample it, the more
> accurately we can time-tag the input condition when it occurs. We'd like to
> obtain an accuracy of better than the sampling interval at 352.8
> ksamples/sec. (Hope I'm not missing something here.)
You might be missing something. It all depends on what "measurement of the
time when the analog waveform satisfies a certain condition" means. If
that waveform is fairly low frequency sampling at an absurdly high
frequency may not give you better accuracy in making your measurement and
could even make it less accurate under certain conditions. OTOH if your
so called waveform is some sort of sudden event like a step or a pulse
that has high frequency components then the higher sampling rate might
make sense.
Without any explanation of the details the answer to your question is
going to be of the form that your elephant is a wall or rope or tree
trunk.
-jim
>
> How much of a processing load is created at this speed still has to be
> determined however. The integer ratio isn't important - was just using it
> to be brief.
Reply by Eric Jacobsen●December 22, 20072007-12-22
On Sat, 22 Dec 2007 11:13:01 -0500, Jerry Avins <jya@ieee.org> wrote:
>Eric Jacobsen wrote:
>
> ...
>
>> There seems to be a disconnect here of some kind, because I know you
>> know that. ;)
>
>Yeah. Actually, the OP wants to sample fast in order to localize certain
>events in time. I told him that merely oversampling won't do more than
>interpolating after the signal is captured, but that using a higher
>cutoff in his anti-alias filter and sampling accordingly would indeed
>help. That's sampling faster, but not oversampling in the usual sense.
>If you can suggest better, I'm sure he would appreciate it.
>
>Jerry
We were responding to different things, then.
Eric Jacobsen
Minister of Algorithms
Abineau Communications
http://www.ericjacobsen.org
Reply by Jerry Avins●December 22, 20072007-12-22
Eric Jacobsen wrote:
...
> There seems to be a disconnect here of some kind, because I know you
> know that. ;)
Yeah. Actually, the OP wants to sample fast in order to localize certain
events in time. I told him that merely oversampling won't do more than
interpolating after the signal is captured, but that using a higher
cutoff in his anti-alias filter and sampling accordingly would indeed
help. That's sampling faster, but not oversampling in the usual sense.
If you can suggest better, I'm sure he would appreciate it.
Jerry
--
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������
Reply by Eric Jacobsen●December 22, 20072007-12-22
On Fri, 21 Dec 2007 16:19:55 -0500, Jerry Avins <jya@ieee.org> wrote:
>Eric Jacobsen wrote:
>> On Fri, 21 Dec 2007 11:37:01 -0500, Jerry Avins <jya@ieee.org> wrote:
>>
>>> Eric Jacobsen wrote:
>>>> On Fri, 21 Dec 2007 09:47:03 -0600, Vladimir Vassilevsky
>>>> <antispam_bogus@hotmail.com> wrote:
>>>>
>>>>> Jerry Avins wrote:
>>>>>
>>>>>> There are commercial 96KHz systems. There are technical reasons that
>>>>>> make it desirable to exceed 48KHz for some processing operations, but
>>>>>> none for reproduction.
>>>>> If you look at the output spectrum of an audio DAC, you will see the
>>>>> huge amount of noise at the frequencies above the sample rate. The
>>>>> fairly decent filter is required to get rid of that. The noise is the
>>>>> artifact of the noise shaping. It can possibly affect the quality due to
>>>>> the nonlinear effects, and it causes the EMC problems, too. But what is
>>>>> more important this noise shows up on the A-curve noise measurements,
>>>>> spoiling the otherwise nice figures of SINAD. So there is some sense in
>>>>> using the higher sample rates.
>>>> That makes pretty good sense, actually. Use a high output sample
>>>> rate so that you have some unused spectrum to which to move the shaped
>>>> noise, then remove that with the reconstruction filter.
>>> I don't get it. The noise before filtering always goes above the sample
>>> rate, and the reconstruction filter is designed to remove it. A
>>> higher-than-needed sample rate allows some noise to be lower than the
>>> sample rate, but it needs to be removed anyway. What's the advantage for
>>> playback?
>>>
>>> Jerry
>>
>> How do you move the shaped noise above the sample rate without it also
>> being within the Nyquist region?
>>
>> As has been discussed, the higher the output sample rate the more room
>> there is spectrally for the shaped noise and filter transition band.
>
>I understand why getting more bits worth of significance needs faster
>sampling -- there's more than one way to make that trade -- but once the
>low-noise signal is acquired, why keep the superfluous samples?
>
>Jerry
I think we're talking about a DAC application here and not an ADC. The
increase in the sample rate before DAC conversion provides some
benefits in noise shaping and reconstruction filter design. That can
be done regardless of the original sample rate, e.g., after
upsampling.
There seems to be a disconnect here of some kind, because I know you
know that. ;)
Eric Jacobsen
Minister of Algorithms
Abineau Communications
http://www.ericjacobsen.org
Reply by George●December 21, 20072007-12-21
Thanks Jerry ... :o)
"Jerry Avins" <jya@ieee.org> wrote in message
news:NMidnZUBbo3Qp_HanZ2dnUVZ_uevnZ2d@rcn.net...
> George wrote:
>> "Jerry Avins" <jya@ieee.org> wrote in message
>> news:qq-dnTTzvceWsPHanZ2dnUVZ_vamnZ2d@rcn.net...
>>
>> <snip>
>>
>>> The faster you go, the more data there is to sift through for patterns.
>>> The oversampling doesn't add more usable data -- the signal is
>>> completely described by the practical rate -- it just adds more samples
>>> to look at and discard. Is an integer ratio of actual to minimum sample
>>> rate a requirement? Why, for example, 8x. Why not 9.16x?
>>
>> True about not adding more usable data. But in this application what
>> we're looking for is accurate measurement of the time when the analog
>> waveform satisfies a certain condition. The more frequently we sample
>> it, the more accurately we can time-tag the input condition when it
>> occurs. We'd like to obtain an accuracy of better than the sampling
>> interval at 352.8 ksamples/sec. (Hope I'm not missing something here.)
>
> That's a processing trade-off then. The extra samples you get will contain
> no more accurate information than you would get from interpolating at the
> lower rate, *provided the higher sample rate really is oversampling*. At
> the higher rate, you can include higher frequencies in the samples stream
> and thereby achieve better time discrimination than you could through the
> anti-alias filter suitable for the lower rate. That is sampling faster,
> but not oversampling as I understand the term.
>
>> How much of a processing load is created at this speed still has to be
>> determined however. The integer ratio isn't important - was just using
>> it to be brief.
>
> Good luck.
>
> Jerry
> --
> Engineering is the art of making what you want from things you can get.
> �����������������������������������������������������������������������
Reply by Jerry Avins●December 21, 20072007-12-21
George wrote:
> "Jerry Avins" <jya@ieee.org> wrote in message
> news:qq-dnTTzvceWsPHanZ2dnUVZ_vamnZ2d@rcn.net...
>
> <snip>
>
>> The faster you go, the more data there is to sift through for patterns.
>> The oversampling doesn't add more usable data -- the signal is completely
>> described by the practical rate -- it just adds more samples to look at
>> and discard. Is an integer ratio of actual to minimum sample rate a
>> requirement? Why, for example, 8x. Why not 9.16x?
>
> True about not adding more usable data. But in this application what we're
> looking for is accurate measurement of the time when the analog waveform
> satisfies a certain condition. The more frequently we sample it, the more
> accurately we can time-tag the input condition when it occurs. We'd like to
> obtain an accuracy of better than the sampling interval at 352.8
> ksamples/sec. (Hope I'm not missing something here.)
That's a processing trade-off then. The extra samples you get will
contain no more accurate information than you would get from
interpolating at the lower rate, *provided the higher sample rate really
is oversampling*. At the higher rate, you can include higher frequencies
in the samples stream and thereby achieve better time discrimination
than you could through the anti-alias filter suitable for the lower
rate. That is sampling faster, but not oversampling as I understand the
term.
> How much of a processing load is created at this speed still has to be
> determined however. The integer ratio isn't important - was just using it
> to be brief.
Good luck.
Jerry
--
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������
Reply by George●December 21, 20072007-12-21
"Jerry Avins" <jya@ieee.org> wrote in message
news:qq-dnTTzvceWsPHanZ2dnUVZ_vamnZ2d@rcn.net...
<snip>
> The faster you go, the more data there is to sift through for patterns.
> The oversampling doesn't add more usable data -- the signal is completely
> described by the practical rate -- it just adds more samples to look at
> and discard. Is an integer ratio of actual to minimum sample rate a
> requirement? Why, for example, 8x. Why not 9.16x?
True about not adding more usable data. But in this application what we're
looking for is accurate measurement of the time when the analog waveform
satisfies a certain condition. The more frequently we sample it, the more
accurately we can time-tag the input condition when it occurs. We'd like to
obtain an accuracy of better than the sampling interval at 352.8
ksamples/sec. (Hope I'm not missing something here.)
How much of a processing load is created at this speed still has to be
determined however. The integer ratio isn't important - was just using it
to be brief.
Reply by Jerry Avins●December 21, 20072007-12-21
Eric Jacobsen wrote:
> On Fri, 21 Dec 2007 11:37:01 -0500, Jerry Avins <jya@ieee.org> wrote:
>
>> Eric Jacobsen wrote:
>>> On Fri, 21 Dec 2007 09:47:03 -0600, Vladimir Vassilevsky
>>> <antispam_bogus@hotmail.com> wrote:
>>>
>>>> Jerry Avins wrote:
>>>>
>>>>> There are commercial 96KHz systems. There are technical reasons that
>>>>> make it desirable to exceed 48KHz for some processing operations, but
>>>>> none for reproduction.
>>>> If you look at the output spectrum of an audio DAC, you will see the
>>>> huge amount of noise at the frequencies above the sample rate. The
>>>> fairly decent filter is required to get rid of that. The noise is the
>>>> artifact of the noise shaping. It can possibly affect the quality due to
>>>> the nonlinear effects, and it causes the EMC problems, too. But what is
>>>> more important this noise shows up on the A-curve noise measurements,
>>>> spoiling the otherwise nice figures of SINAD. So there is some sense in
>>>> using the higher sample rates.
>>> That makes pretty good sense, actually. Use a high output sample
>>> rate so that you have some unused spectrum to which to move the shaped
>>> noise, then remove that with the reconstruction filter.
>> I don't get it. The noise before filtering always goes above the sample
>> rate, and the reconstruction filter is designed to remove it. A
>> higher-than-needed sample rate allows some noise to be lower than the
>> sample rate, but it needs to be removed anyway. What's the advantage for
>> playback?
>>
>> Jerry
>
> How do you move the shaped noise above the sample rate without it also
> being within the Nyquist region?
>
> As has been discussed, the higher the output sample rate the more room
> there is spectrally for the shaped noise and filter transition band.
I understand why getting more bits worth of significance needs faster
sampling -- there's more than one way to make that trade -- but once the
low-noise signal is acquired, why keep the superfluous samples?
Jerry
--
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������
Reply by Jerry Avins●December 21, 20072007-12-21
George wrote:
> "Jerry Avins" <jya@ieee.org> wrote in message
> news:Eamdnb0ukrr5vPHanZ2dnUVZ_h3inZ2d@rcn.net...
>
> <snip>
>
>
>> George,
>>
>> This appears to be a one-off application, so $.35 or $5.35 probably isn't
>> a strong consideration. You could run at a couple of MHz if you want to.
>> Where will you store the data?
>>
>> Jerry
>
>
> Hi Jerry,
>
> You're right - cost isn't a strong consideration.
>
> Let me see if I understand what you mean by "a couple of MHz". With 16-bit
> encoding at 8X, that would be 352.8 kHz sample rate and 5.644 Mbps bit rate.
> You were referring to the sample rate?
>
> A couple of MHz sample rate is about five times faster than the above, which
> comes to 40X if I understand you correctly. That's what I was trying to get
> to.
>
> Regarding data storage, there isn't any. The stream out of the ADC is being
> analyzed on the fly to look for patterns, then thrown away.
The faster you go, the more data there is to sift through for patterns.
The oversampling doesn't add more usable data -- the signal is
completely described by the practical rate -- it just adds more samples
to look at and discard. Is an integer ratio of actual to minimum sample
rate a requirement? Why, for example, 8x. Why not 9.16x?
Jerry
--
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������
Reply by Eric Jacobsen●December 21, 20072007-12-21
On Fri, 21 Dec 2007 11:37:01 -0500, Jerry Avins <jya@ieee.org> wrote:
>Eric Jacobsen wrote:
>> On Fri, 21 Dec 2007 09:47:03 -0600, Vladimir Vassilevsky
>> <antispam_bogus@hotmail.com> wrote:
>>
>>>
>>> Jerry Avins wrote:
>>>
>>>> There are commercial 96KHz systems. There are technical reasons that
>>>> make it desirable to exceed 48KHz for some processing operations, but
>>>> none for reproduction.
>>> If you look at the output spectrum of an audio DAC, you will see the
>>> huge amount of noise at the frequencies above the sample rate. The
>>> fairly decent filter is required to get rid of that. The noise is the
>>> artifact of the noise shaping. It can possibly affect the quality due to
>>> the nonlinear effects, and it causes the EMC problems, too. But what is
>>> more important this noise shows up on the A-curve noise measurements,
>>> spoiling the otherwise nice figures of SINAD. So there is some sense in
>>> using the higher sample rates.
>>
>> That makes pretty good sense, actually. Use a high output sample
>> rate so that you have some unused spectrum to which to move the shaped
>> noise, then remove that with the reconstruction filter.
>
>I don't get it. The noise before filtering always goes above the sample
>rate, and the reconstruction filter is designed to remove it. A
>higher-than-needed sample rate allows some noise to be lower than the
>sample rate, but it needs to be removed anyway. What's the advantage for
>playback?
>
>Jerry
How do you move the shaped noise above the sample rate without it also
being within the Nyquist region?
As has been discussed, the higher the output sample rate the more room
there is spectrally for the shaped noise and filter transition band.
Eric Jacobsen
Minister of Algorithms
Abineau Communications
http://www.ericjacobsen.org