LTC2508 32 bit ADC

[martinr33]

How about this:

http://www.linear.com/product/LTC2508-32

Datasheet
http://cds.linear.com/docs/en/datasheet/250832f.pdf


32-bit ADC. That's - theoretically - good enough for 9 1/2 digits. Plus, it is only $12.

The question is - how far can anybody push this thing? It is one and a bit digits better than anything we have today, it needs 100 picovolt quality designs.

Looks like a fantastic part to let us look into the finest tools we have today,

[zlymex]

One thing to look at is the linearity, which is the one I love the best for 3458A(0.05ppm), and LTC2508 is ten times worse(0.5ppm typ). Another thing to look at is the noise, which is also much inferior than modern 8.5 digits DMM.

However in the past, there are some people making DMM by using this kind of ADC to achieve results rival commercial 6.5 digits DMM or even 7.5 digits.

[David Hess]

The linearity of a delta-sigma converter is limited by its quantiziser which is why I am always skeptical when people suggest dithering an ADC or DAC while expecting greater accuracy.  The first INL Error versus Input Voltage graph shows this nicely with 16 peaks from a major bit transition.

Usually INL error can be calibrated out but that will be much more difficult in this case compared to a converter which uses a single bit quantiziser like the LTC2400 which has a simple INL error curve.  Maybe it could be done with a precision ramp or low distortion sine wave and the 22 bit composite code.

[coppice]

The linearity of a delta-sigma converter is limited by its quantiziser which is why I am always skeptical when people suggest dithering an ADC or DAC while expecting greater accuracy.

I think you see a lot of people talk about dithering for greater accuracy, because most people easily confuse accuracy and resolution.

[TiN]

Btw, seems LT also has unreleased (yet?) LTC2512-24 ADC. It's probably similar to LTC2508 arch, but faster filter and lower resolution, only 24 bits.
That would be nice for precision ~100kHz AC measurements! 32bit one is too slow for that frequency range.

LTC2512-24 mentioned in DC2222A demoboard manual, but LT site has no product page for LTC2512-24 

[David Hess]

Btw, seems LT also has unreleased (yet?) LTC2512-24 ADC. It's probably similar to LTC2508 arch, but faster filter and lower resolution, only 24 bits.
That would be nice for precision ~100kHz AC measurements! 32bit one is too slow for that frequency range.

LTC2512-24 mentioned in DC2222A demoboard manual, but LT site has no product page for LTC2512-24 

The 0.5ppm INL will always be the limiting factor.  Linear Tech also came out with a 20-bit SAR ADC (the LTC2378-20) that also has this 0.5ppm INL limitation, but it will take 1 million readings per second.  So, you can over-sample and average to remove noise, but you will never be able to get beyond this 0.5ppm limitation.

If the INL is stable, it can be calibrated out.  Linear Technology mentions this in a couple of places and an example of blind first and second order correction of an LTC2400 is shown in application note 86 with the mentioned caveat that "The strata is becoming rarified when “error contribution” is delineated in fractional parts-per-million and the yearly drift rate noted".  I wonder though how heroic an effort it would take to do this with a SAR ADC.

AC measurements should be made with a thermal voltage converter [TVC].  There is just nothing that I know of that is commercially available that can outperform a TVC for ability to handle super high crest factors, pretty much any wave shape, and frequencies from DC to 10's of MHz, all while getting 10ppm of accuracy.  Multi-junction thermal converters are available now commercially, and they can get you down to maybe <=1ppm-- so, AC accuracy to 6.5-digits!

Who makes these?

The Datron 1281 [aka: Fluke 8508A] does resistance measurement the "Right Way", and if you are going to copy a resistance measurement circuit, then copy this one.  Better still, would be to build in a DCC for lower resistances [<=10K\$\Omega\$] and an electrometer arrangement (similar to your high-resistance bridge) for everything >=10K\$\Omega\$.  The DCC could be dynamically re-purposed to provide DCI and ACI readings to 100A (again to 1ppm, which is phenomenal).

DCC?

[Kleinstein]

The 0.5 ppm INL is quite good, but not that much better than lower cost slow ADCs like the LTC2400. So it might be an option for something like a battery powered 5-6 digit DMM.
With adjustment or extra calibration points one might be able to reduce the INL, at least the slow varying part of it that is stable.

Using one of the faster LT converters to do RMS conversion in software (just like the 3458, keysight meters call this true volts) is now a really realistic way for high resolution AC measurements. The bandwidth may be a little lower than the analog RMS chips, but it's more accurate and stable. One could even have the option to do filtering in software and thus measure RMS values with intentionally lower bandwidth, which could be interesting for low voltages. The crest factor for thermal converters is also limited by the input amplifier - unless you work without it, with lots or care and a low impedance source.
I would expect the high end digital lock-in amplifiers to use such convert(s).

[tszaboo]

Btw, seems LT also has unreleased (yet?) LTC2512-24 ADC. It's probably similar to LTC2508 arch, but faster filter and lower resolution, only 24 bits.
That would be nice for precision ~100kHz AC measurements! 32bit one is too slow for that frequency range.

LTC2512-24 mentioned in DC2222A demoboard manual, but LT site has no product page for LTC2512-24 

The 0.5ppm INL will always be the limiting factor.  Linear Tech also came out with a 20-bit SAR ADC (the LTC2378-20) that also has this 0.5ppm INL limitation, but it will take 1 million readings per second.  So, you can over-sample and average to remove noise, but you will never be able to get beyond this 0.5ppm limitation.

If the INL is stable, it can be calibrated out.  Linear Technology mentions this in a couple of places and an example of blind first and second order correction of an LTC2400 is shown in application note 86 with the mentioned caveat that "The strata is becoming rarified when “error contribution” is delineated in fractional parts-per-million and the yearly drift rate noted".  I wonder though how heroic an effort it would take to do this with a SAR ADC.

You will drive this from a 5V or 2.5V reference voltage. A degree change in temperature will have several dozen LSB change in the output code. So your calibration table will be changing constantly. With these SAR ADCs you need to be aware that they need a stable, low impedance, high bandwidth (several MHz, really) reference source, and the ADC will have different load to the reference based on the input voltage. Same goes for driving the input. So you cannot really just filter it with RC, you need a fast, high current, not drifting opamp, which cannot really be zero drift, because the chopping frequency is just too low.

[David Hess]

Using one of the faster LT converters to do RMS conversion in software (just like the 3458, keysight meters call this true volts) is now a really realistic way for high resolution AC measurements. The bandwidth may be a little lower than the analog RMS chips, but it's more accurate and stable. One could even have the option to do filtering in software and thus measure RMS values with intentionally lower bandwidth, which could be interesting for low voltages.

Tektronix apparently did exactly this with their handheld DMM series of multimeters which have RMS measurement capability better than their contemporary analog computing counterparts.

Fluke initially sued Tektronix for trademark infringement over the color of the protective rubber sleeve and then after Tektronix changed the color to blue, Fluke bought the entire handheld multimeter business from Tektronix to keep them from competing with their own technically inferior high end products.  I do not know what became of them after that.

... So you cannot really just filter it with RC, you need a fast, high current, not drifting opamp, which cannot really be zero drift, because the chopping frequency is just too low.

Use the chopping amplifier to correct the drift of and remove the 1/f noise from a fast precision low noise amplifier.

[Vgkid]

Who makes these?

Best technologies, another one is made by a japanese company


[quote
DCC?
[/quote]
Direct Current Comparator.
Guildline, Measurements International, Isotech(resistsnce thermetry)

[EmmanuelFaure]

Multi-junction thermal converter data sheet attached.

Pretty nice! An idea of the price range?

[TiN]

DCC means "Direct Current Comparator", and this is used for current ratios in a resistance bridge, or for current ratios when measuring DC (and up to about 100KHz).  See: Danfysik and LEM DCCT's...

I think these results might help someone. Was playing a bit with my LEM Ultrastab IT600-S and calibration equipment while ago.
Also have plan to redo similar test on my home gear, but that will come later on
Brief and non-destructive teardown of these zero-flux current sensors was done year ago and published here.

Ok, here are the results.

Current source 1: Fluke 5700A
Current source 2: Fluke 5700A + 5205A
Configuration : 7 turns thru LEM
Burden shunt: 2.5 ohm Vishay PG foil 1ppm/C 9W, CSNG
Meter: Keithley 2002 on 200mV range
All gear was freshly calibrated and kept within 2C of cal temp.

0A : 0 mVDC rel
214.1942 uA : 0.00250 mV
2.141942 mA : 0.02505 mA
21.41942 mA : 0.249960 mV
200mA : 2.33431 mV
214.1942mA : 2.49986 mV
1A : 11.6719 mV
2A : 23.3434 mV
2.141942A : 25.00005 mV

High current tests with Fluke 5205A (uncalibrated) + 5700A (calibrated)

1.000000 V(MFC) : 11.64732 mV
2.000000 V(MFC) : 23.31058 mV
2.142857 V(MFC) : 24.98688 mV

With 5205A:

3.000000 V : 34.99193 mV
4.000000 V : 46.66403 mV
5.000000 V : 58.33468 mV
6.000000 V : 70.00847 mV
7.000000 V : 81.68046 mV
8.000000 V : 93.35251 mV
9.000000 V : 105.024558 mV
10.000000 V : 116.69671 mV
11.000000 V : 128.36888 mV
12.000000 V : 140.04104 mV
13.000000 V : 151.71314 mV
14.000000 V : 163.37437 mV
15.000000 V : 175.05746 mV
16.000000 V : 186.72914 mV
16.300000 V : 190.23057 mV

5205A could not go higher, as it's not suitable for large inductive loads and was losing stability after 16A.

[EmmanuelFaure]

Very interresting TiN. Maybe we should start a topic on DCC/DCCT?

Some great introduction informations in this document by LEM : http://www.digikey.com/Web%20Export/Supplier%20Content/LEM_398/PDF/lem-engineering-applications-manual.pdf

[b_force]

The chain is as good as the weakest link.
32 bits sounds fun, but getting that accuracy in your resistor divider and reference voltage is gonna be a pain in the butt.
Especially if you involve temperature, noise and component tolerance.

[guenthert]

I'm not complaining, 0.5ppm is very good-- in fact better than the specs for most 6.5-digit and 7.5-digit DMM's.  Others have pointed out that the extra resolution can be useful for relative readings-- well, yes, it can, but maybe only one additional decimal digit.

What?  What makes that one digit special?  INL is a limitation of accuracy, but not precision and there is value in precision without accuracy (well, perhaps not in metrology, but outside of it, particularly if one is interested in trends).

Beyond that is "marketing hype", and that additional decimal digit cannot be relied upon to be absolutely accurate-- (kind of like "deceptive advertising" or an "attractive nuisance").

Unless they claim those 32bit to be accurate, I don't see how it is deceptive.

[guenthert]

I'm not complaining, 0.5ppm is very good-- in fact better than the specs for most 6.5-digit and 7.5-digit DMM's.  Others have pointed out that the extra resolution can be useful for relative readings-- well, yes, it can, but maybe only one additional decimal digit.

What?  What makes that one digit special?  INL is a limitation of accuracy, but not precision and there is value in precision without accuracy (well, perhaps not in metrology, but outside of it, particularly if one is interested in trends).

Beyond that is "marketing hype", and that additional decimal digit cannot be relied upon to be absolutely accurate-- (kind of like "deceptive advertising" or an "attractive nuisance").

Unless they claim those 32bit to be accurate, I don't see how it is deceptive.

I think you are confused.  The INL is a measure of how precise the ADC is throughout the entire range of the ADC.

I might be confused, but it seems I'm not alone in that: Maxim calls INL and DNL parameters of accuracy: https://www.maximintegrated.com/en/app-notes/index.mvp/id/283

[guenthert]

Hmmh, for precision we need monotonicity, but monotonicity is guaranteed only if INL <= 0.5 LSB [1].  Given that, I too, don't know what those extra bits are giving us. 
Update:
Well, strike that.  It says right there that |INL| <= 0.5 LSB is a sufficient, but not necessary criteria for monotonicity.  I.o.w. a larger INL doesn't preclude monotonicity (which strictly speaking isn't a requirement for precision either, but unless one can maintain a large translation table, it rather is a practical requirement).  So I don't know.


[1] http://www-inst.eecs.berkeley.edu/~ee247/fa08/files07/lectures/L12_f08.pdf  (p. 20)

[David Hess]

So, what I'm getting at, is that Linear Tech has a 24-bit version of this same part, and it is just as "precise" as this 32-bit part, because I think you will find that the lower 8 bits of this 32-bit part are nothing but noise and are useless-- even with very long integration times.  I think that those extra 8 bits are just bullshit, for marketing purposes only.  If someone can prove that the lower 8 bits of this part are repeatable and monotonic, then I will admit I'm wrong about this part.

I would also like to know how they tested the 32 bit part however the datasheet unambiguously says that all three outputs, 32, 14, and 8 bit, are monotonic.  The 32 bit INL versus voltage graph shows the typical 1-bit delta sigma converter S shape overlayed with a 4 bit major transition of a higher resolution multibit modulator.

The DC histogram with the maximum decimation factor of 16k shows only 4 codes out of about 1300 samples.  At 61 samples/second, that is 20 seconds which seems almost reasonable in some applications.  With additional processing, it really does look like 32 bit resolution is available with an integration time below that.

I agree that both parts will be equally accurate as far as INL.  I am surprised the INL is competitive with a delta sigma converter using a 1-bit modulator at all.

[Kleinstein]

The limit to get a monotonic curve is  DNL < 1 LSB. Even than if for few events there will steps that are not monotonic any more, a little bit more resolution of the ADC helps, as it could prevent additional rounding errors in the calculations the follow. The often state the resolution of the high resolution converters as the limit where you get no missing code, which would be more of less the DNL < LSB limit.

By some clever method, internal adjustment these SAR ADCs seem to adjust the steps very well. At least the DNL and INL curves look relatively smooth, with very little effect of MSB steps. For the 32 Bit Version the oversampling nature could be some help. The higher speed sigma delta converters also use muli bit DAC/ADC - so no more pure 1 Bit. So this one is more like the other extreme: 24 bit ADC and massive oversampling.

These 32 Bits ADC and the faster 24 Bit ones show some noise. So single conversions are not repeatable to that degree. But after enough averaging (could take more than 100 samples) the average should be OK, provided the reference, signal and temperature are that stable. If you want to compare noise, one should do it at the same rate: for the LTC2508 with 16000 averaging mode one is at about 60 samples per second. The DS gives an RMS noise of 0.02 ppm.  The specs of the 3458 are about 0.08 ppm (RMS=SD) at 1 PLC (or 0.02 ppm at about 20 PLC) in the 10 V range.  To be fair one has to add noise of the input amplifier. Still noise wise this ADC is really good.

So it may be a lower noise alternative to something like the LTC2410 in some applications.

[e61_phil]

Very interresting TiN. Maybe we should start a topic on DCC/DCCT?

Oh yes, please!

[b_force]

As in my previous post, I am still wondering about the practical implematations.
32 bits with a 2.5V supply voltage, means that eacht step is about 0.58nV.
Which is 0.000000023%.
Even if we forget things like temperature drift. Practiaclly it's almost impossible to get in the same order of magnitude with your supply voltage (and reference voltage) . On top of that the input stage and attenuation circuit.
In fact to do it right, you'll need these things an order of magnitude better.

[Andreas]

How can you ask such questions?
Perhaps you accidently are in the wrong section.

> Metrology
> Everything metrology & Calibration related. This is where the Voltnuts hang out.

A real voltnut wants a 64 Bit ADC to play with.

With best regards

Andreas

[David Hess]

As in my previous post, I am still wondering about the practical implematations.
32 bits with a 2.5V supply voltage, means that eacht step is about 0.58nV.
Which is 0.000000023%.
Even if we forget things like temperature drift. Practiaclly it's almost impossible to get in the same order of magnitude with your supply voltage (and reference voltage) . On top of that the input stage and attenuation circuit.
In fact to do it right, you'll need these things an order of magnitude better.

This is less of a problem with ratiometric operation which is likely for applications listed on the datasheet.

[Echo88]

Does anyone know where to get a DC2222A-B Demo-Board which uses the LT2508-32? As far as i can see, the only shop which sells them is Digikey and they have none of them stocked. Also the sales representative couldnt tell me how long it would take Linear to ship the boards.

I just want to use the LT2508 as an ADC to accurately measure thermocouples and play around with the achievable low noise floor of the LT.

[Kleinstein]

For a thermocouple you could use amplification before the ADC, as normally the voltage is small. So no need for the high resolution, just a low noise level. The LTC2508 is relatively low impedance at the input anyway and thus usually needs at least a buffer.

So for a TC it would be better to have a really low noise OP (e.g. ADA4522 or similar)  for something like a 10-100 fold amplification. TC are not that accurate, so no need to worry about absolute accurate amplification. After that the ADC noise is less critical.