Linear Tech Reference Filter Design Note (568)

[Conrad Hoffman]

Thought this was interesting- http://cds.linear.com/docs/en/design-note/dn568f_web.pdf
Any filter is only good to some low frequency cutoff point, so can't fix everything, but this seemed well thought out.

[zhtoor]

very interesting.

if somebody (maybe andreas) could actually measure the sub-10Hz noise, that would be great.

regards.

[Kleinstein]

The filter will not improve low frequency noise. It also gets difficult to set the frequency limit very low, as capacitors have limits. One can simulate / calculated the frequency response rather easy.

Even with a moderate frequency limit, it can take quite some time for the filter to settle to very low levels due to DA effects in the capacitors.

[MisterDiodes]

That circuit has been published for a few years now if you work with LT apps engineering - that's what LT offers as a solution on how to use a '6655 with their "32 bit" ADC's.  Which if you look at the specs the 32-bitters are not really 32 bit for absolute measures and LT will be the first to admit these are not true 32-bit ADC's at DC.  LT/AD doesn't have a single chip, low noise Vref solution for the high resolution ADC's. 

Your ADC noise is only as good as its Vref noise especially as the input signal level approaches Vref level.

Bear in mind that if you using any of these faster, high res ADC's your Vref has to be buffered anyway and must be able to deliver juice fast to the ADC Vref input pin.  You can try hanging a capacitor on the Vref pin with somewhat less than stellar results - you'll get what you pay for.  A quiet amp that can really drive that pin at ADC Vref pin at switch freq works better.

These ADC's have another "gotcha": Their INL is not very predictable like '2400 series either, which makes it nearly impossible to compensate in software.  You do get a faster read rate if you're application needs that but you might be better off looking at some 18 or 20 bit units for lower THD on AC signals. 

As far as the filter circuit - it helps a bit if you're having to use a noisy, drifty '6655 in a ratiometric application but as Conrad pointed out:  You're never getting rid of 1/f noise, and you're not averaging out 1/f noise either, no matter what math you throw at it.  For metrology apps better stick to LTZ for the Vref for less noise in the first place.  Once you start filtering the Vref you keep wishing that it worked for lower and lower freq's.    Then you add more brute force filtering and realize that it takes a long time to settle, and around and around it goes. Good luck with that.

NOTE:  In actual practice TH polystyrene caps work much, much better in the filter.  PP films are a good choice too with low DA.  The downside to those is the relative large real estate required so there's always a trade off.

For us this filter technique and 32-bit ADC's has always been a bit "meh" if you're after quiet, accurate, low drift absolute measures.  Depends on the application.

[Echo88]

Hmm i thought you mentioned that RC-filtering for a LTZ1000 isnt possible metrology-wise without spoiling the LTZ-drift? Or do you refer to the described filter for high speed ADCs instead of metrology applications?

Btw: Page 260 AoE 3: In the footnote it is mentioned that a "Tom Bruins from Agilent" did really long measurements on film capacitors to find that the insulation resistance is significantly higher than stated in datasheets and is commonly measured wrong due to dielectric absorption. I couldnt find any info on Tom Bruins and havent bothered yet to write Mr. Winfield Hill a mail. Maybe some of you guys know where Tom Bruins hides his knowledge?

[Kleinstein]

For a fast ADC filtering makes absolute sense, even with a low noise reference like the LTZ1000. With higher noise references filtering is even more important. The filter shown is with an 1 kHz cross over, thus rather fast. In this frequency range filtering is relatively easy. Already this filter circuits uses a 20 µF cap - quite a size if one goes for PP type to get low DA.

Even with slow integrating ADCs, there might be some sensitivity to reference noise in the 10-100kHz range due to a kind of aliasing. This might might not be that much, but still worth some simple filtering, if the reference is relatively noisy, like the LM399. For the SD ADCs it could be noise in the MHz range that might enter to a small amount.

For precision "DC" measurements one tends to spend 10s to 100s of seconds for a single reading. So the relevant frequency range in more in the mHz (not MHz). In this range analog RC filtering is not practical. The more practical lower limit for a filter is more in the 1-10 Hz range. At some point a better reference gets easier.

[MisterDiodes]

If you're using a noisier '6655 you'll probably want filtering on the faster ADC's especially, If you're using an LTZ on then faster ADC's you'll probably want a low-noise driver for the ADC pin after your 7V / X volts divider.

The other point I was making was to be careful about attaching a zero-drift chopper amp  - directly - to an active die substrate without giving the input chopper switch spikes somewhere to go, otherwise that's not a good idea to subject something lie the LTZ die to those spikes.  Dr. Frank and Andrea's have circuits with '2057 amps but they do have bypass caps on the '2057 inputs.  The advice from my experience and  directly from LT Apps engineering is that they really don't want the '2057 used as the -current driver- in the datasheet circuit (without paying attention to the input current noise spikes) and really the '1013 is always the best amp to use for an LTZ current driver for lowest current noise on the die - that is the amp perfectly designed and tailored for the LTZ.  Just about any amp will still work for the current driver but you want lowest injected noise on the LTZ die for best long term stability as that gives the crystal lattice the best chance at aging gracefully and quietly over long time frames.

Again:  Voffset of a '1013 current driver has really no big effect on the Vref output because of the transistor gain on the LTZ, so there is no real reason to use a zero-drift  amp for the LTZ current driver.  An output buffer is a different application.

Using a zero-drift amp as part of an output buffer will work - just pay careful attention to switching spikes on the inputs and on power rails.  We always isolate the LTZ die from the chopper buffer inputs with a very modest RC and have no problems, Dr. Frank has a circuit here that just uses a C...so do whatever works for you.  But you do have to pay attention to those switching currents.  It is best to look at how zero drift / chopper amps work, and realize that mosfet gate switching current is going to add noise to your input signal at the amp's switching frequency.  This part is usually not found in the datasheet, you have to measure it - and once you do it is surprising.  The input current noise listed in the datasheet at low freq and the current noise at chopper freq are two VERY different things.

The circuit shown as an example from LT is one way to build a low-drift amp with better, faster output drive current - usually the zero-drift amps will have very low drive current capability.

[maxwell3e10]

I've been mulling a way to make a very low-noise voltage reference with 1/f corner frequency below 1 Hz. Here is one idea: use a very low-pass filter with a large mF capacitor but add a PI feedback to servo it to a stable voltage reference. If the filter capacitor is well thermally insulated, then any changes in its voltage due to temperature changes can be compensated by feedback.  This also assumes that the load on the output changes very slowly. I have to still do a simulation but I am wondering if anyone has seen such a circuit or can see an obvious problem with it.

Another possibility maybe is to use a NiCd battery, which is known to have very low noise, and a feedback to charge it slowly.

[KE5FX]

Hmm i thought you mentioned that RC-filtering for a LTZ1000 isnt possible metrology-wise without spoiling the LTZ-drift? Or do you refer to the described filter for high speed ADCs instead of metrology applications?

Btw: Page 260 AoE 3: In the footnote it is mentioned that a "Tom Bruins from Agilent" did really long measurements on film capacitors to find that the insulation resistance is significantly higher than stated in datasheets and is commonly measured wrong due to dielectric absorption. I couldnt find any info on Tom Bruins and havent bothered yet to write Mr. Winfield Hill a mail. Maybe some of you guys know where Tom Bruins hides his knowledge?

Tom's name is spelled 'Bruhns', if I remember correctly.

[Echo88]

Thanks! Tom Bruhns leads to some fitting search results.

[tszaboo]

I've been mulling a way to make a very low-noise voltage reference with 1/f corner frequency below 1 Hz. Here is one idea: use a very low-pass filter with a large mF capacitor but add a PI feedback to servo it to a stable voltage reference. If the filter capacitor is well thermally insulated, then any changes in its voltage due to temperature changes can be compensated by feedback.  This also assumes that the load on the output changes very slowly. I have to still do a simulation but I am wondering if anyone has seen such a circuit or can see an obvious problem with it.

Another possibility maybe is to use a NiCd battery, which is known to have very low noise, and a feedback to charge it slowly.

Doesnt work. Your cap will generate noise.
Also, stick a 8.5 digit meter on a battery, and try measuring the voltage. Good. Now, squeeze the battery with your fingers.

Which if you look at the specs the 32-bitters are not really 32 bit for absolute measures and LT will be the first to admit these are not true 32-bit ADC's at DC.

Datasheet states, that it is a 14 bit SAR ADC with digital magic. Now LT has a very nice SAR ADC architechture, with 0.5PPM INL, but the 20 bitters will offer better performance in general. Also, these shouldn't be compared with the Sigma-deltas. These will create a sample every microsecond or so, not every second.

Dr. Frank and Andrea's have circuits with '2057 amps but they do have bypass caps on the '2057 inputs.

I'm having trouble understanding, why they would use an auto-zero amp in this circuit. It looks to me that marketing told them to sell that opamp. Something like a LT1028 would be a much-much better choise there. The 0.8uV/K drift is 0.2 ppm/K. The input-output noise figures are a lot-lot better, and it has 50MHz GBW, so possibly it could drive the ADC input.
Just compare the two diagrams, 2057 G17 and 1028 G08. The LT1028 is using 10nV/div while the 2057 is 100nV. I'm sorry, but I'm not impressed by this app note. Also, there are no danish cookie boxes or paint cans in it.

[blackdog]

Hi NANDBlog,

Dit you look at the bias current of a LT1028?

Kind regards,
Blackdog

[tszaboo]

Hi NANDBlog,

Dit you look at the bias current of a LT1028?

Kind regards,
Blackdog

No I havent. I spent 5 minutes choosing an opamp which is potentially better than what they used, and I havent wrote an article. Bias current will matter much less than you think. Sure, it makes a few uV of offset, yes. But it is stable. According to 1028 G14 from 25 Celsius to 50, the bias only changes some 3-4nA. Through that 10Kohm resistor it is 30-40uV. That is about 30uV/25K on 5V, or 0.2-0.3 ppm/K. Besides, if you choose that opamp, the architecture of the filter is different.
Your milage may vary.

[Conrad Hoffman]

I try to take advantage of the fact that reactances have no Johnson noise whenever I can. Certainly a large capacitor and long time constant circuit will have noise/drift problems due to the inevitable temperature change, but other than that, I don't understand why there would be any noise source in a cap. Also, I've found batteries are not as much of a low noise device as people (especially audiophiles) seem to think.

[Kleinstein]

The input bias can matter quite a lot. The input bias is not only slightly temperature dependent, but it can also change over time. Besides the voltage noise, there is also significant  current noise with BJT based OP. With a 10 K source the LT1028 is not at all low noise  - more like about on par with the LTC2057 and with extra 1/f noise. For an 10 K resistor a good BJT based OP would be something like the LT1001 - still low noise, but not that low.

Anyway the noise of even the low noise references is relatively high compared to many OPs. The LTC6655-5 or LM399 are at about 100 nV/Sqrt(Hz) - that is about 10 times the noise voltage of the LTC2057. In addition the LM399 has significant 1/f noise. Even the LTZ1000 is higher in noise than the LTC2057, LT1013 or similar OPs at most frequencies.

At higher frequencies, the typical LTZ circuit will also include the noise from the OP, as the transistor is only active for the low frequencies. So if one would care about the higher frequency (e.g. 10 kHz range) noise, it would help a little (the zener noise is still a little higher) to have a lower noise OP in the LTZ1000 circuit. Here however filtering would be an alternative for this frequency range too.

The fast high performance ADCs can have input noise levels that are lower than most refs. So filtering makes absolute sense if one cares about the kHz range. In the Hz range using a lower noise ref or several refs in parallel might be the better option. The usual way is combining with resistors and this can be combined with a cap to get higher frequency filtering too.

[tszaboo]

Does it really matter? As I said, I'm not impressed by this appnote. It looks like someone put it together the output of a filter designer, and TI's slyt355 appnote.
If they want to impress me with a low frequency filter, recreate the Datron 4910 7 pole PWM filter circuit with modern parts. Surprisingly, that also has below 1Hz cutoff frequency, and it is very low noise. But it is so forgotten, the topology doesnt even have a name.

[chuckb]

The Datron 4910 Filter reminds me of a Frequency Dependent Negative Resistance (FDNR) filter. I first saw this style used to clean up the output of digital audio DACs at 44 kHz (see attached TI app note). It's probably used in every digital audio device out there.

Then you AC couple the filter to not disturb the DC signal. I first saw that in LT AN20. The LT App Note is to large to attach here.

[bopcph]

That circuit has been published for a few years now if you work with LT apps engineering - that's what LT offers as a solution on how to use a '6655 with their "32 bit" ADC's.  Which if you look at the specs the 32-bitters are not really 32 bit for absolute measures and LT will be the first to admit these are not true 32-bit ADC's at DC.  LT/AD doesn't have a single chip, low noise Vref solution for the high resolution ADC's. 

Admitted - the LTC2508-32 is really not a fair "32bitter" as it starts with 14bit diff. SAR and adds an 8bit common mode.
If you want some serious "near 32bits" you should start with the LTC2500-32 - it samples 24bit diff. SAR and adds 7bit common mode.
It performs 28bits as stable as the HP3458A with the same LTZ1000A as Vref - only it does it 5 to 10 times faster for the same ENOB.
(refering to HP3458A spec.sheet)

On paper the LTC2500-32 will do conv. of a +/-5V input (5Vref) to 32 bits at 61Sps with a noise floor of 0.2uVrms (using SinC avg. on-chip) - though the resolution 1 LSB would represent 2.33nV 

These ADC's have another "gotcha": Their INL is not very predictable like '2400 series either, which makes it nearly impossible to compensate in software.  You do get a faster read rate if you're application needs that but you might be better off looking at some 18 or 20 bit units for lower THD on AC signals. 

The software in - was it LT AN81 ?, A precision voltage reference" !? - actually does a fair job compensating for the LTC2400 INL of 4ppm and the nonlinearity of the LTC1599 

Doing sampled and calculated AC measurements I would go for a much higher sample rate than both the 24' and 2500 series can do.
AC measurements with a "resolution" of "140dB" is not for the faint hearted  - just see how we struggle to do our noise measurements - relative AC measurement are much easier

The amplifier - putting the LT6202, with 100MHz GBWP - low noise and fair gain, inside the FB-loop of a LTC2057 makes perfect sense.
The 62.. got what it takes to make a low noise filter and the 2057 keeps all the DC parameter in check.

As the LTC2508 is sampling at 1MHz and the update rate at the output is 61sps, you do not need a filter that can "shave" the noise down to 0.1Hz.

[GigaJoe]

That interesting,

I'm scratching my head , trying to recall , about capacitor implementation   .... Specifically requirement for C3
And  found:  https://www.embeddedrelated.com/showarticle/742.php

I'm I right? That article above seems missing one significant point, about upper capacitor, as leakage related to the applied potential, the upper capacitor has a significantly higher impedance then lower one.