new voltage reference from MAXIM

[Andreas]

Hello,

similar game with some LT1027CCN8-5 (DIP plastic package) specced 3 ppm/K with datasheet method.

Temperature variation of samples #4 to #7 ranges from 60uV to 250uV over a 35 deg C temperature range.
Giving 0.5-2.1 ppm/K average T.C.
Hysteresis is much lower than with LTC6655 (usually below +/- 1ppm).
With some "cold creeping" effects (I blame it to the plastic package).

ADC8 shows the regression curve of sample #3.
The resulting error is around +/- 2uV or +/-0.5 ppm.

So I am curious how the LS8 package will perform.

With best regards

Andreas

Edit: of course the temperature compensation will not help against the humidity sensitivity of the plastic DIP package which is around 0.5ppm/% rH for the LT1027C. This will give around 15 ppm seasonal change (40-70%rH) here in Germany.

[uncle_bob]

Hello,

similar game with some LT1027CCN8-5 (DIP plastic package) specced 3 ppm/K with datasheet method.

Temperature variation of samples #4 to #7 ranges from 60uV to 250uV over a 35 deg C temperature range.
Giving 0.5-2.1 ppm/K average T.C.
Hysteresis is much lower than with LTC6655 (usually below +/- 1ppm).
With some "cold creeping" effects (I blame it to the plastic package).

ADC8 shows the regression curve of sample #3.
The resulting error is around +/- 2uV or +/-0.5 ppm.

So I am curious how the LS8 package will perform.

With best regards

Andreas

Hi

Numbers 5 and 6 may be exhibiting a bit of dependance on the low end temperature. it also could be a number of other things.

Bob

[Kleinstein]

The hysteresis in temperature cycles can also include some humidity effect: heating lowers relative humidity levels and thus drives out water from the plastic of the case and the board. It can take quite long (weeks and more) for humidity to come back at room temperature.

The temperature compensation / numerical correction only helps a little with low TC parts, but it can be very helpful with parts with a higher TC. So with compensation the main parameters to look at are long term drift and hysteresis - the TC is not that critical any more. If measured in the real circuit and with the typical slow varying temperature, the thermal coupling is not critical and also linearity of the temperature sensor is no critical, it's mainly repeatability and long term stability that is important. So a temperature sensor in a plastic case (like many of the digital chips) is not such a good idea,as they can show similar hysteresis.

[tszaboo]

The problem with this is that it is very hard to predict behaviour too. How strong is thermal coupling between thermistor and referenc, what exactly is the compensation required, each reference requires manual adjustment as coefficients are varying...
So, does not replace a tempeature stabilized (oven) solution.

Someone finally got it right!
You can characterize a reference with a heat chamber and it is possible to compensate it's change with a different part. But the characterization is very expensive, time consuming, labor intensive and so on. It is not really an option for a production environment, while it might be an option for a one-off reference.
I did calibration of reference voltages. With a 3458A. It was a major operation, with extra hardware designed to burn in stabilize, and measure the references multiple times to get the confidence level. It was a week to do some few hundred boards calibrated. 30 minutes warm up, 1000 NPLC, several measurement, and paralleling the work. It was just barely possible to do this, for references, which deserved it (the one with the 1000 in its name).Now, include a temperature chamber test, selecting the right compensation component soldering, burn-in again, verification... It is just not possible. And not worth it, because once you need to spend labor to do it, mind as well use a heated reference.

[uncle_bob]

The problem with this is that it is very hard to predict behaviour too. How strong is thermal coupling between thermistor and referenc, what exactly is the compensation required, each reference requires manual adjustment as coefficients are varying...
So, does not replace a tempeature stabilized (oven) solution.

Someone finally got it right!
You can characterize a reference with a heat chamber and it is possible to compensate it's change with a different part. But the characterization is very expensive, time consuming, labor intensive and so on. It is not really an option for a production environment, while it might be an option for a one-off reference.
I did calibration of reference voltages. With a 3458A. It was a major operation, with extra hardware designed to burn in stabilize, and measure the references multiple times to get the confidence level. It was a week to do some few hundred boards calibrated. 30 minutes warm up, 1000 NPLC, several measurement, and paralleling the work. It was just barely possible to do this, for references, which deserved it (the one with the 1000 in its name).Now, include a temperature chamber test, selecting the right compensation component soldering, burn-in again, verification... It is just not possible. And not worth it, because once you need to spend labor to do it, mind as well use a heated reference.

Hi

Doing 100% temperature test and characterization may not be production compatible in all areas, but it is in some. We run a *lot* of parts through temp comp every month. Some of the stuff sells for < $20 in volume.

Bob

[uncle_bob]

I think you could automate this with the right setup.  Perhaps running up to 100 devices per 24-hour run in an environmental chamber.  Building the automated DAS and acquiring the necessary instrumentation and low thermal EMF switches might be rather expensive, so you would only do this if you are running many production runs.  As far as I can tell, there just is not much of a market for this kind of thing.  After [and *IF*] you sell your first hundred units, you will be hard pressed to sell another 100 I think.  To make any money from this, there has to be a "Pot O' Gold at the end of the rainbow"...  It could be that the massive investment in money, equipment and time will end up being a huge loss.  UNLESS you are a charlatan, and are just installing an epoxy packaged reference IC inside a plastic box, and then selling it to unsuspecting dupes on eBay...

If you are just doing this as a hobby, well then the amount of money you spend on the reference and the amount of "labor" you put into the project is irrelevant, right?

There are some very good references in epoxy packages that are good enough for 4.5 digits of accuracy [100ppm], and there are some monolithic ICs in hermetic packages that might be good enough for 5.5 digits of accuracy [10ppm], but if you want or need accuracy >= 6.5 digits [1ppm], then you are going to have to go with a heated reference, and at this level of accuracy, it's hard to beat the LM399 for lower cost projects or the higher performance LTZ1000(A) for projects where you can afford the best.

All of this talk about getting super high performance [1ppm] out of an unheated monolithic reference IC is becoming tiresome, because it has been explained over and over in multiple threads why this will never happen.  Any effort to this will just be a "fool's errand"-- so, don't be a fool.  If it sounds too good to be true, then there is almost a 100% certainty that it is false.  This fact of life also applies to voltage references, and also to charlatans on eBay claiming to provide you with 2ppm of accuracy for $99...

Hi

The thing you need to do is to scale your test setup to the market demand. If you figure that the total available market is a hundred a year, a single channel system is probably plenty good enough. Maybe two single channel setups if you worry about this sort of thing a lot. The cost of running the chamber might change that a little. You don't need a very big test volume for a single piece, so maybe not.

Bob

[splin]

For under a dollar, you can get a *bunch* of different solid state temp sensors. They have different accuracy ratings, but in general are all very repeatable and stable long term.
Bob

It would be interesting if you could quanitfy long term stability as it is rarely specified for low cost sensors (and often for expensive devices too) and if it is the spread is usually large - eg. LMT70 specifies drift as < 10mK typical, +/= 100mK max, 10k hours @ 90C. Presumably you used calibrated sensors to monitor their stability?

[uncle_bob]

For under a dollar, you can get a *bunch* of different solid state temp sensors. They have different accuracy ratings, but in general are all very repeatable and stable long term.
Bob

It would be interesting if you could quanitfy long term stability as it is rarely specified for low cost sensors (and often for expensive devices too) and if it is the spread is usually large - eg. LMT70 specifies drift as < 10mK typical, +/= 100mK max, 10k hours @ 90C. Presumably you used calibrated sensors to monitor their stability?

Hi

We plot them against stuff that should show us 10s of mK over the course of months. That testing is done on the entire system rather than an individual component. Unless the sensor *always* happens to drift in the same direction as the rest of the system .... we would see it.

Bob

[splin]

Hmm, I'm not quite sure what you mean - you're comparing against a number of different sensors/types? If you aren't taking any absolute, calibrated measurments surely you can only conclude that some (perhaps most) drift at similar rates and any that don't stand out beause they are different - but those could be the ones that aren't drifting significantly?

Isn't this the problem of having lots of clocks or voltage references and never being sure which are right?

[uncle_bob]

Hmm, I'm not quite sure what you mean - you're comparing against a number of different sensors/types? If you aren't taking any absolute, calibrated measurments surely you can only conclude that some (perhaps most) drift at similar rates and any that don't stand out beause they are different - but those could be the ones that aren't drifting significantly?

Isn't this the problem of having lots of clocks or voltage references and never being sure which are right?

Hi

Not really:

You have a voltage reference, it is out in the ambient environment
You have a voltage controlled gizmo that you are compensating, it has a known sensitivity to the control voltage
You have a temperature sensor that gives you the data for the compensation process
You compensate the whole thing to a few 10's of ppt using the temperature data from the sensor
You validate the compensation process and put the beast on long term burn in
You come back in 3 months / 6 months / 12 months and see how everything is holding up
If the compensation holds to a few ppt or so of what it was, you can back into the stability of the parts.

The compensation of the device is a temp sensor driven process so the end result is the sum of all the errors along the way. You can put upper bounds on things, but not lower bounds. We have a few points in the system we can monitor and take out a couple things, it's still an "upper bound only" estimate.

Yes, there are a bunch of grubby details about sensitivity and how tightly you are compensating things vs how bad they were at the start. Since it's my employers "secret" process,I can't get into quite everything. I suspect I've put enough in there that you can pretty well guess what we are playing with.

Bob

[Kleinstein]

Having a heated reference does more than just reduce the TC. It also reduces hysteresis quite a lot, as the thermal history will be similar after each turn on. Also the relative humidity of the heated part will be low - so the humidity effect is also reduced. However there are two down sides: the part gets hotter and thus ages more, and the power consumption is higher, usually by something like a factor of 3 at least.

[lars]

Any experience with this LT1021CMH supply?

http://www.ebay.co.uk/itm/191123710347?

I've bought a couple and plan to test these but right now I have very little "free" time to do it. It looks good however and the price is OK for a NOS metal case C grade reference.

Cheers

Alex

Hi,
 
I bought two in March last year.

In December I put them in boxes with NTC temperature compensation ala SVR-T. I did the same with 2pcs of REF102CM and 6pcs AD587 in ceramic packages I had in my junk box. So totally ten boxes. The boxes are 35x55x20mm plastic and the binding posts are gold plated brass from China. The 15v power is a 5.5/2.1mm connector and I also have two small 4mm banana jacks connected to a 10k NTC. The PCB is just a prototype board. So a lot of not optimal solutions.

For the LT1021CMH I used two 470k NTC (SMD0805) in parallell as I had that in my junk box. I added a resistor about the same value in series to linearize the NTC as I didn´t bother to compensate the second order component. To trim the temperature coefficient I used a trimpot from the  junk box between the output and ground. Used another trimpot and resistor to trim the voltage.

The LT1021CMH´s had +5 resp. +7ppm/C before compensation (well within spec of 20ppm/C). They were about 200ppm low.

Enclose a chart for the first 6 months of test with continuous power. The refs probably was powered down a couple of weeks in end of March as I had an intermittent connection on the +15V supply. On April, 1 I had them powered 4 hours before measuring. All values are relative to a SVR-T board. As null meter I have used a homemade x100 amplifier + 3/2 digit DMM to get 0.1ppm resolution.

The result for the LT1021 are better than I hoped. 5 years ago I bought 6pcs LT1031DMH from Distrilec/ELFA and was very disappointed. See attached graphs. During two periods I have turned of power between the measurements and only turned on it for 1hour for the 2pcs in room temperature and one period for the 4pcs in a box about 13C above room temperature. As can be seen it is a terrible difference with power on continuous instead of off for several weeks between power on. I will do this test with the LT1021 later. as it seems so different from the LT1031.

Lars

[lars]

Some comments about temperature compensation or stabilization (not inside the IC) of voltage references:

In my opinion it takes time both to design and test the design for both approaches. Design is probably easier with an NTC compensation ala SVR-T. For professional work I prefer to go with stabilization as production test time is expensive. Trimming a compensated hobbyist reference takes time but not necessarily advanced equipment. More than 15 years ago I trimmed the compensation on my first REF102CM´s with NTC´s by just warming them with a lamp (at that time it wasn´t a LED lamp . Three out of four has about 0.0ppm/C at 23C and the fourth I must have made a mistake as it is 0.4ppm/C

For compensation I have used NTC´s for designs with compensations if no uC is involved. NTC´s are non-linear and that is perfect for both first and second order compensation over a narrow temperature range (as in the SVR-T). In my professional work I very seldom design in NTC´s but instead silicon sensors. I really have no idea how the silicon sensors drift but I guess well below 0.1C per year. Assume you want to compensate a volt ref with 5ppm/C. 0.1C per year drift will give 0.5ppm/year. Not easy for most of us to measure.

Stabilization might not be better than compensation as you still need to consider placement of the sensor and temperature gradients and how they change with surrounding temperatures. A more complex design might also have internal changing gradients due to varying power.

I also would like to say that my hobbyist goals are quite different from my professional. My goal as a hobbyist has always been simple and cheap designs. In my professional work I get enough of complicated and expensive equipment (most >1MUSD each). Being a volt nut for just 20+ years but a curious hobbyist for at least the double time I would really encourage to experiment with voltage references if you like measurements and calculations as I do.

Another comment: As a professional I almost 100% go by the data sheet spec but as a hobbyist it is nice to characterize the voltage ref IC´s myself and see if I can get better performance.

Lars

[Gyro]

Hi Lars,

Very interesting data, thank you. Can I ask, were your AD587 samples already aged? They seem very stable if not.

It looks as if I can have good confidence in the long term drift of my SVR-T. 

Chris

[lars]

Yes, the AD587 and REF102 was salvaged from old boards.

But still I think you can have very good confidence in the SVR-T. I have four SVR-T and four SVR boards converted to SVR-T. They are of different ages but still within a couple of ppm from each other and I have compared them to one of my REF102CM´s that has been powered for more than 15 years and the drift is less than 2ppm compared over about 5 years for my oldest SVR.

I also have two AD587LQ datecode 0045 that I got from Joe five years ago. They have been powered and in a box with about 13C above room temperature (30-43C). One has drifted +3ppm rel the REF102CM and the other -3ppm. During three months 2014 I turned off the box they are in and only powered up them for three hours about every third week. During that time I measured after 1 and 3hours that can be seen in the attach graph. Still the graph is almost a straight line. Compare to the LT1031´s! in the same box.

These two AD587 are not temperature compensated but instead four LM35´s are in the box and data post processed. They have a tempco of about +2ppm/C

My best guess for the SVR-T is about 5ppm expanded uncertainty as a transfer std (3 weeks) and 10ppm over 5 years.

Lars

[Gyro]

Many thanks Lars, more data than I could possibly have hoped for. 

I think I remember Joe only having acquired one big batch of AD587LQs so they are probably all 'immediate family', mine is also 0045 (batch C67337).

That's quite a reference farm you're running there!

Chris

P.S. If those REF102s were also salvaged then it potentially also says something about their relative response to physical disturbance.... or maybe even CERDIP vs TO-99?

[HighVoltage]

I just came across this one:
Analog Devices voltage references P/N: AD2710LN
Ebay listing for US $218.00

http://www.ebay.com/itm/201420682682

What is so special, that it is so expensive?
Does anyone here had experience with this reference.

[lars]

As null meter I have used a homemade x100 amplifier + 3/2 digit DMM to get 0.1ppm resolution.

Hi Lars

Anything special about this amplifier? In this particular application, the offset and offset drift are a bigger issue than gain error and gain drift. Anything special about offset error compensation?

No it isn´t anything special with this amplifier. Had to check my schematic from 1992.
I use a TLC2654 zero drift op-amp and 10k+1Mohm 1% metalfilm. As you say gain error and drift isn´t a big problem, but I may have selected the resistors.
I have no offset compensation. A note from 1992 says -4uV offset and now it is -7uV. As voltage references have low impedance outputs I read the amplifier shorted before measurements.

Lars

[tszaboo]

So last week TI introduced the REF60xx references. They have a high speed buffer built in, otherwise its like the REF50xx family. It is actually really nice. I always had to buffer the REF5025, which was going to those high speed 18 bit SAR ADCs.
BTW if you havent, then get one of linear technology's 20 bit SAR ADCs to play with. Its can measure one PPM, and do it 100.000+ times a second. Truly amazing.

[SoundTech-LG]

REF60xx references...

Nice, I had just ordered, and received the Maxim, but these appear to add some value with the buffer.

[tszaboo]

No, this 24 bit part has some digital filtering in it. I'm talking about the LTC2376-20 - 20-Bit, 250ksps, Low Power SAR ADC with 0.5ppm INL part. I had the chance to send it againtst a Agilent 3458A, and auto calibrate it with a buried zener reference voltage. That was something truely amazing analog frontend.

[uncle_bob]

Hi

I believe you will find that a lot of the 24 bit ADC's have a nasty 1/F noise corner on them. That makes the whole "average to take out the noise" approach a very long and slow process. If you look at the data on the LTC2376-20 there certainly is a low end roll up on all the plots they show. The solution is often to go to a chopper based front end. When they do that, it is up in bold print at the top of the data sheet. They charge you a bit more for it ....

Bob

[Kleinstein]

These LTC237... ADCs use a sampling input stage, so I would not expect much 1/f noise. The input is much like an AZ OP: the very low zero drift (7 ppb/K  which would be something like 30-70 nV/K) indicates this. The main source of 1/f noise if likely the reference voltage source.

The relatively high speed could be an advantage for a chopper type input stage, as there is not much time lost - it could be more like the auto zero mode of a DMM, doing the demodulation in digital and not analog. I don't think this would be needed because of the ADC's LF noise, but more due to noise / drift of the input amplifier itself. For a single reading the noise might not look that low, but it already a challenge finding a suitable reference, especially if you need the LF range too. After averaging down to a lower sampling rate, this gets very competitive compared to DMMs - they too may be limited by the reference.

[uncle_bob]

These LTC237... ADCs use a sampling input stage, so I would not expect much 1/f noise. .....

Hi

That's what I though before I actually had to make it all work. Turns out that there are still very real noise sources in the signal path.  People like TI make parts with choppers in front of them.  Yes, I did go back and yack at Linear about it. They ultimately admitted that there is a 1/F corner on the parts.

Bob

[tszaboo]

These LTC237... ADCs use a sampling input stage, so I would not expect much 1/f noise. .....

Hi

That's what I though before I actually had to make it all work. Turns out that there are still very real noise sources in the signal path.  People like TI make parts with choppers in front of them.  Yes, I did go back and yack at Linear about it. They ultimately admitted that there is a 1/F corner on the parts.

Bob

Just look at the "DC histogram" figure. So that is the amount of noise you will get. Getting rid of the pink noise is real pain. Ultimately you need to background calibrate the signal chain (zero measurement on any DMM) and use averaging/longer sampling. ie a 8.5 digit DMM will only give you true 8.5 digit result, if you use very long NPLC settings (1000NPLC on a 3458A). So turn up those integration times. Also mains frequency synchronization might be really necessary.
And yes, choppers help, and they dont help. They increase the wideband noise above the chopping frequency. So you need to filter it. For example with an RC, increasing the impedance driving the ADC. So you buffer it... with a chopper?