32-bit ADC playground for precision measurement tasks.

[The Soulman]

±0.5ppm INL (Typ) and ±2ppm INL (Min/Max) is not that impressive. Even the good old LTC2400 delivers 4ppm INL and 0.3ppm RMS noise but with a predictable and therefor adjustable error. Thus the error can be adjusted to <1ppm. LTC2500-32 seem to have not that easy adjustable INL error.
But maybe TiN can prove us that I'm wrong.

What's more revealing is if you convert those figures into counts (~4300 and ~17000 respectively) and think about what this really means in the context of a "32-bit" ADC. (Hint: log₂(4300) = 12 and log₂(17000) = 14).

4294967296?  Huge resolution but poor accuracy?

[MK]

For an actual piece of equipment you'd use in the lab:  A older piece of equipment that is calibrated, working, documented and well-aged is even more priceless and useful.

If you need an extra digit on your DMM...   is it realistic to try to build a very high quality x10 amplifier to put in front of a good 5 or 6 digit meter?

Then you need an accurate source to generate a matching voltage to greater than 5 or 6 digit stability to get an extra digit on a voltage measurement...

[David Hess]

The slight difficulty with this ADC and the LTZ1000 (and other 7 V references) is that one would likely need to divide down to something like 1/3 or 1/2. This divider would also need to be very stable, like a capacitive or maybe a precision transformer circuit.

I was thinking of a switched capacitor voltage divider.  Didn't Keithley use them for linearity calibration of their ADCs?

[tszaboo]

This produces that ~20 bit noise free bits every 10 us (if it is anything like the 2376, that I've thoroughly tested)...

Once you actually use a 250x part, you're not really going to see that, honestly. Maybe 18 or 19 bits.  Probably like 2376.  But your Vref is modulating your input data, and at low ppm's that becomes more and more critical. As your signal increases dynamic range then the Vref noise becomes even more critical.  Then if you're running these 250x fast with an FPGA good luck keeping the that digital noise out of the faster measure (if you're chasing PPMs)...Uggh. It can be done, but you have to be careful.

Speaking of noise: Also notice that the AC specs (SNR, THD) of the '250x parts are pretty lackluster.  That's more like a 16 or 18 bitter.

An '2400 / 2408 has 4ppm INL but it is normally corrected out in software, and has been since the 80's. Linear even gives you the code for it, it's not like you have to re-invent the wheel.  You should get to ppm absolute or sub PPM relative measures on every part in a production run.  Very repeatable and stable correction curve is very similar even across parts - and is stable over time (decades).  We just use a 10 point test cal correction during manufacture and it's quick and easy.  You can get to ppm (or 20 bits) relatively easily, and if you keep your system very very quiet you can get into the 21~22 bit range.  No good way to correct a '2500 / '2508 long term, they seem to bounce around every 1kHrs or so.  We've been looking at them for the past year or so and feeding suggestions back to LT, but they seem a bit confused these days.  This seems to be a part that marketing wanted to have a datasheet for to compete with AD, but now they are the same company.  The will probably drop either the LT or the AD 32-bit part at some point, no need for two parts competeing against each other.

Again:  These '2500 serires are not meant to be a 32 bit, sub-PPM absolute measure device, and LinearT is the first to tell you that.  Use these when you need to measure a high-resolution (not necessarily with high absolute accuracy) ratiometric sensor.  Use a '2400 / '2404 / '2408 with a 6655 or LTZ's for accurate DC measuring at a MUCH better profit margin.

If you check the histograms on page 8 of the datasheet, at DF = 64 we have 1ppm noise (AKA about 20 bit) DF=64 is 16 KSPS. VREF noise is indeed critical. But. Anything below 10 Hz can be compensated by the secondary reference (you just throw away some of the throughput), and anything above 10Hz can be filtered simply by RC filtering and then buffering the reference.

Well of course nobody is talking about sub PPM measurements. Anything below 0.5 ppm is lunacy, since even the 3458A has 0.5ppm/24h drift on its best range. I dont know about you, but it would be a little ridiculous for me, to just casually throw a Fluke 732A on a schematic next to some resistors and ADCs.
With off the shelf parts, getting 20 bit noise free, it is relatively easy. I can create an AFE, which performs better than a 6.5 digit meter on a single range. If the range is current (100mA+), than the task is even easier. Calibrating it, in house is also cheap to some 20-30 ppm, although, I would never do 10 point calibration. The bottom line is: Today's SAR ADCs are better than yesterday's delta-sigma.

I have a hunch, that TiN will try to push these ADC a little harder than that. I never had that specification. Although once i spent days proving that 0.01 ppm system is not possible. Turns out later, a manager did not know, that ppm does not mean "parts per mille" (0.1%) but "parts per million". That was fun. I glanced at the specs, said outright "That is not possible". But you know, when you are manager, you dont need to listen to engineers anymore.

What's more revealing is if you convert those figures into counts (~4300 and ~17000 respectively) and think about what this really means in the context of a "32-bit" ADC. (Hint: log₂(4300) = 12 and log₂(17000) = 14).

I dont know, where those numbers are coming from, but they are magnitudes off. 4ppm is 250.000 and it is about 16 18 bit.

[The Soulman]

is 250.000 and it is about 16 bit.

With 16 times oversampling?

[Marco]

The slight difficulty with this ADC and the LTZ1000 (and other 7 V references) is that one would likely need to divide down to something like 1/3 or 1/2. This divider would also need to be very stable, like a capacitive or maybe a precision transformer circuit.

It only has to be stable in between auto-calibration cycles ... which shouldn't be a very long time.

[Cerebus]

What's more revealing is if you convert those figures into counts (~4300 and ~17000 respectively) and think about what this really means in the context of a "32-bit" ADC. (Hint: log₂(4300) = 12 and log₂(17000) = 14).

I dont know, where those numbers are coming from, but they are magnitudes off. 4ppm is 250.000 and it is about 16 bit.

?


Full_Count *  4 ppm = (2³² - 1) * 4 / 1000000 = 17179.869... counts

(250000/(2³² - 1)) * 1000000 = 58.2076... ppm

log₂(250000) = 17.93... bits

Either I've lost it entirely, or you need a new calculator.

[tszaboo]

What's more revealing is if you convert those figures into counts (~4300 and ~17000 respectively) and think about what this really means in the context of a "32-bit" ADC. (Hint: log₂(4300) = 12 and log₂(17000) = 14).

I dont know, where those numbers are coming from, but they are magnitudes off. 4ppm is 250.000 and it is about 16 bit.

?


Full_Count *  4 ppm = (2³² - 1) * 4 / 1000000 = 17179.869... counts

(250000/(2³² - 1)) * 1000000 = 58.2076... ppm

log₂(250000) = 17.93... bits

Either I've lost it entirely, or you need a new calculator.

First, 250.000 is indeed 18 bit. DOH!
And I realized, you were calculating the number of bits "lost".

The slight difficulty with this ADC and the LTZ1000 (and other 7 V references) is that one would likely need to divide down to something like 1/3 or 1/2. This divider would also need to be very stable, like a capacitive or maybe a precision transformer circuit.

It only has to be stable in between auto-calibration cycles ... which shouldn't be a very long time.

Exactly, if you do the calibration more often than 10 hz, you calibrate out most of the 0.1-10hz noise.

[SilverSolder]

I overlaid some bit depths on a chart of theoretical measurement limits (from a Keithley document) to get a sense of how crazy it is to try to measure DC with 32 bits of resolution.  (Referenced to 1V)

With a sufficiently low source impedance, it doesn't seem completely insane?

[Spikee]

Can't you just simply divide the signal in an upper half and a bottom half and measure those with two 20-24 bit adc's to
get a higher overall resolution? close or higher than 28 bit ?

[Kleinstein]

The 32 Bit resolution is more like the data format. I would guess they just did not consider any value between 24 Bits and 32 Bits. Restricting it to just 24 Bits could add a little to the errors, so they to the next 8 Bit step. However some of the 24 Bit SD adc's also give 30/32 Bit data without calling them 32 Bit - but the noise limit is usually at less than 24 Bits.

The INL of some of the SD converters, like the LTC2400 or LTC2442 is relatively predictable - so there is a chance to numerical compensate for some of it. This essentially does not work with the LTC2500. Still the specified maximum INL for the LTC2500 is about the best one can get from a of the shelf chip ADC (the chip used in the Prema DMMs might be better - but availability is a problem). However after adjustment / selction some of the SD converters might get better values.

The reference could be in deed important to get ultimate possible noise level of the ADC. If not used in a ratio-metric application (e.g. resistive sensor). This could be a problem. For the higher frequencies (e.g. > 10..1000 Hz) one could use filtering. But for the lower frequencies it would help to have a reference with low 1/f noise. Using a secondary reference and thus measure signal and reference in fast sequence would only make sense if this is also needed to avoid 1/f noise in the amplifier chain. It would also take some time and thus increase the overall noise. So it would be a real advantage to really have the reference from a low 1/f noise source and do connection to a secondary reference only on a very long time scale (like hours). I don't think it is practical to use a secondary reference to also get the 0.01 .. 1 Hz range from that reference - it would degrade the overall noise too much, as it would take a considerable part (like 25%) of the time and add the noise of the ref. measurement.

For the datasheets 4 of the LTC6655 seem to be about the same level of 1/f noise as the LTZ1000 at 4 mA. Not sure how real world performance would be. The LTZ circuit would need the extra divider, which could also include higher frequency filtering. Not sure how much noise a capacitive divider would add.

[SilverSolder]

The 32 Bit resolution is more like the data format.

I guess the bit count determines the smallest signal you can read, while overlaid on a larger signal?

So for example, measuring a 1 uV change on a 10V DC signal requires more A/D bits than measuring 1uV "on its own"?

[Andreas]

For the datasheets 4 of the LTC6655 seem to be about the same level of 1/f noise as the LTZ1000 at 4 mA. Not sure how real world performance would be.

You have to be very carefully with the noise specs of the LTC6655.
I don´t know how they really measure them.
(perhaps with some kind of foam to thermally isolate the reference connected only with thin wires.)

On a real PCB (with cotton pads on both sides of PCB),
 I measure for a 5V device between 2.2uVpp and 3.6uVpp which is factor 2-3 above the datasheet value.
Interestingly the variation depends largely on the power supply voltage.
The higher the self heating - the more noise.

but who has a 5.5V rail in a system when he wants to measure up to 10V.

with best regards

Andreas

[Andreas]

The 32 Bit resolution is more like the data format.

I guess the bit count determines the smallest signal you can read, while overlaid on a larger signal?

So for example, measuring a 1 uV change on a 10V DC signal requires more A/D bits than measuring 1uV "on its own"?

Usually the definitions are as follows:

Resolution: (effective resolution) = ld2(full range / rms noise) in bits
so resolution has nothing to do with bit count only with the rms voltage of the noise.
E.g. the LTC2400 has 1.5uV rms noise (around 10uVpp).
With 5V full range you have a resolution of 21.6 Bits.

Bit count: number of valid bits in result register (with no missing codes).
so DNL has to be less than +/- 0.5 LSB to determine the number of bits.
E.g. the LTC2400 has a 28 bit result register but only 24 bits without missing codes. -> 24 Bit converter.

with best regards

Andreas

[SilverSolder]

Resolution: (effective resolution) = ld2(full range / rms noise) in bits [...]
E.g. the LTC2400 has 1.5uV rms noise (around 10uVpp).
With 5V full range you have a resolution of 21.6 Bits.

Is the noise of the ADC due to its input resistance (i.e. the resistor noise of its own input circuitry) or is there something else that dominates?

[MisterDiodes]

A few more thoughts and Head's Up:

 Andreas is correct - most of the time the LTC6655 is more noisy than you're led to believe.  These always show up on the Linear Tech Demo boards when an LTZ-level device would be better suited.  And then you get the 20-bit '5791 DAC demo board from AD and it has the layout for the LTZ for best performance <Grin>.  That tells you everything you need to know on what you're going to need to get to a stable 20 bit accuracy level...

RE: High Rez ADC's.  Talk to Apps engineering at Linear Tech.  Yes the datasheet says 32 bit, and the internal SAR setup might be 32 bit but one thing is guaranteed:  That's never going to be verified at that full resolution absolute - there is nothing to verify against.  They will admit that also.

Modern SAR better than older Delta Sig??  I could be that when you're looking at slightly quieter data it seems that way - you can over sample, dither, gather a lot more samples in an effort to improve ENOB (Effective Number of Bits) and the real red flag is when you see noise figures on datasheets specified at zero signal input.  Get very suspicious at that point.

The reality is that NO amount of math voodoo averaging, SINC filtering or any other bullshit "computer science data gymnastics" is going to conjour up better -absolute accuracy- than what the basic system hardware design provides.  Yes you will maybe get a little quieter data, and yes maybe you can get another bit or two ENOB and get the flicker down a little - but that doesn't necessarily mean you've gotten any better accuracy.

A lot of times we'll take a noisy 16-bit accurate (but very accurate) data over "20 bits flicker free" garbage ADC data any day, no matter how flicker-free it looks.  It depends on the application, and what the customer is willing to pay for, and what true real accuracy is needed.  It doesn't help that the "flicker free" data is sitting nice and quiet several 10's of counts north or south of where it should be, especially when you realize you got some digital switching noise beating in with your data or auto zero amp somewhere.  Oops.  You never know until you really test and verify your input system across all possible input ranges.

In the real world, once you start to go after higher ADC resolutions you notice that Reality is shooting down the Theory in your head right between the eyes when:  a) That Vref is never quite good enough over longer time spans b) That front end circuit feeding that magical ADC is never quite good and stable enough over time and temperature c) Those '5400 resistor paks you put on the front end amps (because LT  datasheets told you to) aren't good enough once you start looking at noise and long term stability d) The voodoo math SINC / averaging / oversample  isn't delivering true better absolute accuracy performance like you thought e) Trying to add a correction curve to a faster SAR chip is probably a fool's errand once you realize it's not stable over time and temperature f) Putting a hot, noisy, expensive FPGA chip right next to that high res ADC is a really dumb idea when you're trying to retrieve a delicate signal g) Once you get the system running and realize that the 32-bit ADC has fairly lackluster stability over time and temperature (because the heat from that expensive, hot and noisy FPGA is causing data trouble), and so on.

Be very careful and consider all the pitfalls.  I've hinted at just a few to get you started.

So no, in my experience newer SAR's are not always a real improvement over a stable and proven '240x series, not by a long shot - at least not in every application.  Now if you do need the sample speed, then the SAR's can do that...but not always the best choice if you're trying to chase down into low PPM's absolute measures.  We had a good experience using a '2500 reading fast (4kHz), faint air pressure changes on a very low air pressure sensor, and using a '6655 as Vref - but this was a ratiometric application.  And these 32-bitters certainly work well for an application like that, as that is the intend market application.  As LT will tell you.

[tszaboo]

For the datasheets 4 of the LTC6655 seem to be about the same level of 1/f noise as the LTZ1000 at 4 mA. Not sure how real world performance would be.

You have to be very carefully with the noise specs of the LTC6655.
I don´t know how they really measure them.
(perhaps with some kind of foam to thermally isolate the reference connected only with thin wires.)

http://cds.linear.com/docs/en/application-note/an124f.pdf
775 Nanovolt Noise Measurement for A Low Noise Voltage Reference
by Jim Williams
Happy reading!

Resolution: (effective resolution) = ld2(full range / rms noise) in bits [...]
E.g. the LTC2400 has 1.5uV rms noise (around 10uVpp).
With 5V full range you have a resolution of 21.6 Bits.

Is the noise of the ADC due to its input resistance (i.e. the resistor noise of its own input circuitry) or is there something else that dominates?

A SAR ADC is a switched capacitor ADC. It means, it switches a (randomly charged) internal capacitor for the voltage reference. For a short time, like nanoseconds in 1MSPS case. In that nanoseconds, you need to charge the internal capacitor to the voltage of the voltage reference. Very accurately, I suggest calculating with 0.5 of your desired accuracy. Easy task at 16 bit, you might get away with just a big 10 uF ceramic capacitor, but it is difficult at 20+ bit and difficult if your samplerate is high.

[MisterDiodes]

NANDBlog - RE: Jim Williams and '6655 - Yes, that's what the datasheets show, and that's how they were built when JWiiliams was around....but have you checked a recent shipment of 6655's?  We just had to return a batch of 2.5V to LT that were an order of magnitude out of whack for noise.  LT did replace them but you have to keep checking.  It seems like something is going on at LT after the takeover.

That's another thing to watch out for with newer SAR's - What you thought was an OK level of noise on your 10V Vref, maybe a few uV...all of the sudden that uV noise becomes about a 4X bigger problem with a 2.5V Vref....

[Echo88]

Reminds me of the modification i had to make on my LT2508-demo board to measure real bipolar signals: add a LT5400-resistor network and a LT6363. Calculating the error resulting from that modification gets you out of those 32-bit and super high accuracy dreams. 

[MisterDiodes]

  Yep - This is definitely the world where Theory Ends and Reality Begins....  You find out in a hurry the only real datasheet is the one you get from building and testing real circuits, for your application.

[SilverSolder]

That's another thing to watch out for with newer SAR's - What you thought was an OK level of noise on your 10V Vref, maybe a few uV...all of the sudden that uV noise becomes about a 4X bigger problem with a 2.5V Vref....

Maybe reference noise could usefully be defined the same way Andreas said above for an ADC...

i.e.  (effective resolution) = ld2(full range / rms noise) in bits 

For a reference, the "full range" is just its voltage.

[Andreas]

http://cds.linear.com/docs/en/application-note/an124f.pdf
775 Nanovolt Noise Measurement for A Low Noise Voltage Reference
by Jim Williams
Happy reading!

Hello,

as we have already discussed in this forum:
the bandwith of the AN124 is lesser than 0.1 .. 10 Hz
lower bandwidth -> lower measured noise.

E.g. for the high passes:
1) 1300uF * 1200R gives 0.1Hz lower bandwith.
2) 165uF * 10K = 0.1 Hz lower bandwidth.
3) root sum square correction = 0.1 Hz lower bandwidth
All series gives around 0.17 Hz lower bandwidth  (or -9 dB instead of -3 dB at 0.1 Hz) and so on.

Sorry thats not a good example for a well designed cirquit.

with best regards

Andreas

[Echo88]

What about AN159: http://cds.linear.com/docs/en/application-note/an159fa.pdf ?

[Cerebus]

http://cds.linear.com/docs/en/application-note/an124f.pdf
775 Nanovolt Noise Measurement for A Low Noise Voltage Reference
by Jim Williams
Happy reading!

[snip]
Sorry thats not a good example for a well designed cirquit.

I don't think Nandblog was offering it as an exemplar of good design, it was a direct answer to your question:

You have to be very carefully with the noise specs of the LTC6655.
I don´t know how they really measure them.

The app note documents the measurement technique and circuit used for measuring the LTC6655 noise for the datasheet.

[tszaboo]

http://cds.linear.com/docs/en/application-note/an124f.pdf
775 Nanovolt Noise Measurement for A Low Noise Voltage Reference
by Jim Williams
Happy reading!

Hello,

as we have already discussed in this forum:
the bandwith of the AN124 is lesser than 0.1 .. 10 Hz
lower bandwidth -> lower measured noise.

E.g. for the high passes:
1) 1300uF * 1200R gives 0.1Hz lower bandwith.
2) 165uF * 10K = 0.1 Hz lower bandwidth.
3) root sum square correction = 0.1 Hz lower bandwidth
All series gives around 0.17 Hz lower bandwidth  (or -9 dB instead of -3 dB at 0.1 Hz) and so on.

Sorry thats not a good example for a well designed cirquit.

with best regards

Andreas

Yes it is a direct answer. The DUT is the LTC6655
I think it is intentional:
" Figure 6, taken at the circuit’s oscilloscope output, shows 160nV 0.1Hz to 10Hz noise in a 10 second
 ample window. Because noise adds in rootsum-square fashion, this represents about a 2% error in the LTC 6655’s expected 775nV noise fi gure. This term is accounted for by placing Figure 3’s “root-sum-square correction” switch in the appropriate position during reference testing. The resultant 2% gain attenuation fi rst order corrects LTC6655 output noise reading for the circuit’s 160nV  noise fl oor
 contribution."
He is mentioning that the caps need 24h because the dielectric absorption. So I guess it took him long time to do all this. But in fact, re-reading it, I've noticed this:
"Figure 8 is LTC6655 noise after the indicated 24-hour
dielectric absorption soak time. Noise is within 775nV
peak-to-peak in this 10 second sample window with
the root-sum-square correction enabled. The verifi ed,
extremely low circuit noise floor makes it highly likely
this data is valid."
So he is saying that in a single 10s sample the noise is less than the datasheet value, with root sum square correction?  I'm seeing some 700nV peak to peak noise on figure 8. And he did strip chart recording to verify that the circuit is not noisy, but there are no strip charts for the LTC6655???
Maybe he is telling us that the datasheet values are wrong? I mean most of the application note is "hey look, my circuit works!".

OK, I looked up the qoute at the end.

"It ceased; yet still the sails made on
A pleasant noise till noon,
A noise like of a hidden brook
In the leafy month of June,
That to the sleeping woods all night
Singeth a quiet tune."

  This is too amusing. Maybe we should forget about that 0.25ppm noise.