Andreas,

Thanks for the information.  Do you have the drift chart, etc. posted
anywhere?  that would be very interesting reading.

Thanks,

Randy

On Fri, Jul 18, 2014 at 10:11 PM, Andreas Jahn Andreas_-_Jahn@t-online.de
wrote:

Hello Randy,

some information you will get on eevblog. (its much easier to post
(larger) pictures there).
Namely within the LTZ1000, LM399 and T.C. Measurements threads:

http://www.eevblog.com/forum/projects/ultra-precision-reference-ltz1000/
http://www.eevblog.com/forum/projects/lm399-based-10-v-reference/
http://www.eevblog.com/forum/projects/t-c-measurements-on-precision-resistors/
http://www.eevblog.com/forum/projects/oshw-24bit-adc-measurement-system-for-voltage-references/
http://www.eevblog.com/forum/projects/building-a-7-decade-voltage-calibrator/
http://www.eevblog.com/forum/projects/ppmgeek!-5-5-digit-dvm-volt-ref-cal-%28for-arduino-or-any-uc-w-spi%29/msg296127/#msg296127

With best regards

Andreas

Am 19.07.2014 16:33, schrieb Randy Evans:

Andreas,

Since you are familiar with the LTC1043, do you know what stability over
time and temperature one could expect for the X2 circuit, assuming a high
quality low leakage capacitor was used?  Since the circuit does not appear
to be sensitive to the capacitor value, the primary change over time and
temperature would be expected to be the switch resistance I would think.
If the circuit were fed into a high impedance buffer (LTC1151), then I
would expect the switch resistance to have minimal impact.  What do you
think?

thanks,

Randy

On Sat, Jul 19, 2014 at 8:58 AM, Andreas Jahn Andreas_-_Jahn@t-online.de
wrote:

Hello Randy,

I have no experience with a voltage doubler.
From theory the output impedance gets halved by a 2:1 divider
and gets doubled with a X2 multiplier.
So I would expect around a factor 4 more influence of error sources.
But this is only a best guess.

The leakage current of the buffer is usually not the limit.
You have to pay attention on the surface resistance of your PCB
and the leakage of your capacitors.

With best regards

Andreas

Am 21.07.2014 18:38, schrieb Randy Evans:

Andreas,

Of course, I would also expect the leakage currents to change over
time/temperature and I would expect them to be the dominant error source.

Randy

On Mon, Jul 21, 2014 at 9:38 AM, Randy Evans randyevans2688@gmail.com
wrote:

On 07/17/2014 10:26 AM, Randy Evans wrote:

Hi Randy,

There are some other error sources that might need to be considered when using
the LTC1043.

I have not used the LTC1043, but note that on the data sheet there is a small
charge injection at each of the switch pins. In the multiply by 2 circuit shown
on the data sheet they are using 1 ufd caps.  Typical charge injection (depends
on voltage level) is 8 pC. With the 1 ufd caps this is 8 uV.  I assume there is
some offsetting effect - but this might be a significant contributor to the 5
ppm error that is mentioned.

There is also a 6 nA (typical) leakage mentioned.  During the "hold" time (
about 1 msec) of the output 1 ufd cap this comes to 6 uV.

Regards,
Bob Smither

yes, charge injection is an issue with all these switches and these also vary, in other words are somewhat unpredictable. now some of this may be compensated by a bigger c, but there are natural limits too. so for a production unit to sell, this would probably be a killer, but also for home-use, how do you predict the behavior over time?

another option to use if you want to stay away from resistors is the pwm-solution as implemented by datron in e.g. their 4910. the pwm signal can today of course be done by uCs.

and yes, I forgot: only down-dividing of course, so to reach 10V, two LTZ1000 would be needed in series. advantage is that noise statistically is reduced by factor of about 1.4. formally also applies to drift.

I agree that there are potentially some serious unknown issues with drift
due to time and temperature due to changes in leakage current, charge
injection, etc.  I would think some serious characterization would be
needed before this approach could be used. One approach for the charge
injection is to try and have the input voltage near Vdd/2 so the charge
injection effects are nulled out.  I'm trying to figure out how to do that
for a X2 circuit.  Any ideas?

I talked to Vishay and they, of course, could do the resistors and I am
awaiting a quote.  They stated they would have to "tune" the vhd200
resistor pair to have a very low (0.2ppm/C) ratio stability .  Doable but
potentially very expensive.  Unfortunately, the maximum value of the
resistors is 20Kohms which would greatly increase the power dissipation of
the resistors, not good for long term stability.  Still working the issue.

Randy

On Tue, Jul 22, 2014 at 8:12 AM, acbern@gmx.de wrote:

Randy,

Have you considered using multiple identical resistors to reduce the
variance? Depending on who you believe, you can reduce the variance of
the overall resistance by SQRT(N) where N is the number of resistors in
series/parallel. Its not that easy to create a good search query for
this but here is one such explanation:

http://paulorenato.com/joomla/index.php?option=com_content&view=article&id=109:combining-resistors-to-improve-tolerance&catid=4:projects&Itemid=4

Ideally they should all come from the same batch - ie. manufactured by
the same machine from the same batch of materials. Obviously there's no
way to guarantee that without close liaison with the manufacturer (you
did want 10 million parts at $.10 each didn't you!) but hopefully a set
of resistors which come off the same reel would come close.

The absolute value isn't important however, but 'statistical gain' will
also apply to the TCR and stability of the overall divider. The
following assumes that both factors are similarly improved by SQRT(N),
but in fact they may be rather better than that.

That80€ or $108 for one sealed Vishay foil divider will buy a lot of
lower spec parts:

Approx 12558 x Susumu RR0510P .5%, 25ppm 0402 (Digikey, $86/10k). 6279
in series and parallel in each leg of the 1:1
dividerhttp://media.digikey.com/photos/Susumu%20Photos/RR%200402%20SERIES.jpg
might reduce the variance to 25ppm/SQRT(6279) = .32ppm. Can't see any
spec for stability, but it may also improve similarly. Would take a
while to solder them onto stripboard though!

Slightly more sensible might be 1078 x TE Connectivity RP73 1%, 10ppm
1206 (Digikey, $100.18/1K).  Stability .5% (no qualifers in datasheet)
=> 10ppm/SQRT(539) = .43ppm, stability => 215ppm

Or 372 x KOA Speer RN731JTTD4021B5 .1%, 5ppm (Mouser, $29/100).
Stability not on data sheet but typical endurance is +/- .02% for 1000
hrs @ 70C on/off 1.5hours/.5hours.
=> 5ppm/SQRT(138) = .37ppm, endurance => 14.7ppm (Stability should be
rather better than that). Note that the Mouser part no. is for a 25ppm
part but their manufacturer's part number is the 5ppm part as is the
description. Also, the price is way too high for 25ppm parts.

Or 28 x Susumu RG2012L .01%, 2ppm (Digikey, $39.6/10). Stability not
quoted but typical Load Life is .01% (1000 x 1.5hours on/.5hours off at 85C)
=> 2ppm/SQRT(14) = .53ppm, endurance => 27ppm

You could also use multiple resistor networks. Eg:

104 x Susumu RM2012B-103/103-PBVW10 .1%, 5ppm tracking, 2
resistors/device (Digikey $104/100). Stability not quoted, endurance
500ppm (1000 x 1.5hours on/.5hours off at 85C)
=> 5ppm/SQRT(104) = .49ppm, endurance => 49ppm

35 x TT Electronics SFN08B4701CBQLF7, .25%, 5ppm tracking 7
resistors/device (Digikey, $76/25) . Stability not quoted, high
temperature exposure < 1000ppm
=> 5ppm/SQRT(122) = .52ppm

33 x TT Electronics 668A1001DLF .5%, 5ppm tracking 8resistors/device
(Digikey, $82/25). Stability not quoted, load life < 1000ppm
=> 5ppm/SQRT(33 * 4) = .45ppm

16 x Vishay DFN .1%, 3ppm tracking with 4 resistors/device (Digikey,
$5.24/1). Shelf life ratio stability is specced at 20ppm (1 year at
25C). (That may be a typical rather than a maximum - your parts may all
be much worse than typical). The 3ppm tracking TCR may also be a typical
figure as its headlined in a section titled 'TYPICAL PERFORMANCE' but in
the specification table its not qualified with '(typical)' as they
sometimes do in other datasheets. Its hard to tell.
=> 3ppm/SQRT(32) = .53ppm shelf life stability => 3.5ppm

5 x Vishay DSMZ metal foil dividers, .5ppm tracking max (probably
performs rather better than this over restricted temperature range, but
don't believe the Vishay typical figure of < .1ppm/C) (Digikey,
$22.93/1). Shelf life ratio stability not quoted but 'typical limit' for
Load Life ratio stability is 50ppm (2000 hours at 70C). Who knows what a
typical limit is? Again, probably best to treat Vishay 'typical' figures
with a pinch of salt given the experience of another poster on volt-nuts.
=> .5ppm/SQRT(5) = .22ppm, load life => 22ppm

Interestingly Digikey quote a price of only $5400 for 1k parts for the
similar DSM divider (1ppm tracking), which is a huge difference from
$22.93. Might be worth considering a bulk buy if there enough volt-nuts
with the same problem. They aren't stocked though so that price might
not be 'real'. However:
20 x Vishay DSM dividers, 1ppm (Digikey, $5400/1000) Load life ratio
stability 'typical limit' 50ppm
=> 1ppm/SQRT(20) = .22ppm, load life => 11ppm

Multiple LT5400 networks could also be used and may give the best
results, but the much larger absolute tolerance, +/-15% would cause
those with the highest value for series connected/lowest for parallel to
dominate and reduce the statistical improvement. Do your own calculations.

Its interesting that all these different components end up providing
pretty much the same performance for the same cost - in other words the
cost is inversely proportional to the TCR^2

My gut feeling is that the tracking TCR will improve rather better than
the SQRT(N) calculated, if they do indeed come from the same batch, as I
would expect them to have similar absolute TCRs. Thus you might be able
to get away with rather less parts to achieve < 1ppm. The SQRT(N) factor
comes from assuming that the variation in the value is random, and I
believe, has a particular distribution (Guassian or normal?). Component
specifications are often derived from the distribution parameters
measured from a large set of production samples, with the max/min values
determined from a multiple (typically 6?) of the standard deviations of
the distribution? The worst case specifications for TCR and stability
may (I don't know, just hypothesizing) be derived very differently. For
example, the TCR may be affected not only by the characteristics of the
bulk resistive material, but also due to stresses on the element due to
thermal expansion of the substrate/packaging. It may be that the former
is almost identical for all components from the batch, but the latter is
less predictable. The specification max/min would have to allow for the
worst cases which might be due to a  relatively few which for some
reason (microcracking in the substrate perhaps) have much larger
variance from the majority. The distribution of TCRs from a set of
resistors could be very skewed with long tails and the SQRT(N) reduction
in variance may be well off the mark.

Stability is more difficult because the shelf life stability is rarely
specified, but is likely to be the closest to your usage. For reference,
the Vishay DFSMZ datasheet specifies ratio stability of .015% for 2000
hour at 70C and .002% for shelf life ratio stability. The 7.5X
difference might be useful for estimating shelf life stability for
resistors that only quote load life or endurance specs. But it might
not! I'm not sure that the endurance spec is very useful either as it
subjects the resistor to a large number of large temperature cycles
which won't be anywhere near your usage.

I would expect the long term tracking stability to be much better than
(worst case datasheet stability)/SQRT(N) as I would expect the vast
majority to age in similar ways, if not by the same magnitude. Whilst
the specs show stability to be +/- xx% I would expect that most will age
in the same way - probably slowly increasing resistance over time. I
also expect there are experienced posters here who know otherwise!
Similarly to TCR, it could be that for example, the stability of most
resistors in a batch may be quite good, but the specs reflect that a few
may be much worse due to random faults in individual samples - such as
defects in the protective coating of the element allowing corrosion to
occur in a few samples. You'd need a very good understanding of the
factors that determine the resistor stability to calculate the overall
stability of multiple resistors.

I would expect similar factors to apply to ratio tracking due to
humidity changes. No doubt there is some useful information out their in
application notes/research papers on the variance in long term stability
between resistors of various types (and maybe even for parts taken from
the same batch) just waiting for some interested volt-nut to discover?

The fewer the parts, the more chance of statistical outliers reducing
the improvement over a single part, but you could test each divider for
the best matching, if you've got a decent meter, fairly easily by
applying a voltage from a stable, low noise source (a battery would be
good if its temperature is kept very stable), and measure the voltage at
the centre tap. Then put the resistor network in a plastic bag  and
immerse it in boiling water to raise the temperature by 75C or so; .5ppm
tracking would give 9.4uV/V maximum change; you'd probably need to
reverse the meter leads a few times to null out thermal EMFs.
Alternatively measure the voltage difference between the divider under
test and another driven by the same voltage source and kept at a stable
temperature - ie. in a bridge configuration. A simple high gain
amplifier (say 1000x) with adjustable offset would allow testing with a
more realistic lower temperature difference of say 20C and/or a cheap meter.

Accuracy is not particularly important - you probably don't need to know
the temperature tracking coefficient to better than 20%.

Component layout would need to ensure any thermal gradients apply
equally to both legs of the divider by interleaving upper and lower
resistors.

Tony H

On 17/07/2014 16:26, Randy Evans wrote: