[volt-nuts] LM399 Long term drift specification

Tony vnuts at toneh.demon.co.uk
Fri Sep 12 12:46:03 EDT 2014

On 12/09/2014 06:16, Andreas Jahn wrote:
> Am 12.09.2014 um 01:06 schrieb Tony Holt:
>>> All ageing specs are "typical" if you want to have "guaranteed" 
>>> values you will have to measure it over a reasonable time. (I 
>>> recommend min 6 months).
>>> Every treatment (soldering, mechanical/temperature shock) of a 
>>> reference may create a new ageing cycle with different slope.
>> True. I guess that the new ageing cycle from soldering in an LM399 is 
>> not going to be as bad as that for a surface mount plastic device.
> Thats true humidity within plastic housing is changed largely by 
> soldering. This gives more stress on the die than on hermetically 
> devices.
> But why are the instrument manufacturers using sockets for the LM399 
> (HP34401A) or reference boards for the LTZ1000 for their pre selected 
> parts?

Well its a lot cheaper to burn-in a device or small reference board than 
the whole instrument PCB which is why I expect to do the same. The LM399 
on the Keithley 2015 isn't socketed though.

>>> So 100ppm/15 years outside of "lab conditions" (23 deg , constant 
>>> humidity) is something that I would not guarantee without 
>>> re-calibration.
>> I had a feeling that would be the answer - though surely humidity 
>> shouldn't be a factor as these are hermetic parts. The questions 
>> remains though, what level might you specify - if you were forced to 
>> come up with a number (ok a guess!) - for non-selected, non-pre-aged 
>> parts after 15years continuous operation without re-calibration? 
>> Obviously this is given the context of the presumably limited numbers 
>> of samples you've tested and I guess you wouldn't have bothered to 
>> further test early rejects.
> Sorry this may change from lot to lot. From other (non heated) 
> references I have very different results. Especially from devices 
> bought before and after ROHS. And partly even better graded devices 
> (LT1027BCN) behave worse than standard grade devices (LT1027CCN) in my 
> temperature range.
> With humidity and LM399: this should be no issue for the metal can 
> package of the LM399. Although I have one LM399 (#1) which I 
> desoldered from a board of unknown age that has a correlation to 
> humidity (see attachment). Ageing is also in the range of 12-15 
> ppm/year average with seasonal modulation which follows humidity. But 
> since this behavior is not typical for my other LM399 references I 
> would justify this sample as "defect".

Fair enough. I guess only the reference manufacturers are going to have 
any worthwhile long term drift data on statistically large numbers of 
devices - apart from customers who aren't going to share the data.

> Humidity: even hermetically parts can suffer from humidity. The epoxy 
> board swells if exposed to humidity. This creates stress to the leads 
> of the package and then to the die. I measure around 0.5ppm/% for 
> plastic parts. In a early publication LT specced around 12ppm change 
> for a 30% rH change for the LT1236AILS8 hermetically package. This 
> spec is now removed from newer data sheets. You will only find a hint 
> to avoid board stress.
Yikes! 12ppm for only 30% rh change in a hermetic package?
>>> Although typical drift of pre-aged + selected references will be in 
>>> the 1-2ppm/year range if properly treated.
>> What would you classify as pre-aged? Do they need to be powered up or 
>> can they be maintained at a suitable temperature? How many rejects 
>> would you expect to get to get one that achieves 1-2ppm?
>> Is it known if the major instrument manufacturers preselect and 
>> burn-in LM399s themselves for their middle-range instruments? I'm 
>> pretty sure the top end kit will be all use carefully tested and 
>> selected parts, but what about a 34401A for example? The basic 
>> accuracy spec for that is 20ppm for 90 days, 35ppm for 12 months so 
>> even a 20ppm guaranteed part wouldn't be good enough, especially 
>> allowing margin for drift in other components. I guess I just 
>> answered my own question!
> From Keithley cal lab you can see that they adjust the instrument 
> during calibration if they fall outside the 70% window. So I guess 
> that after 1 year they think that ageing of the components (including 
> the reference) is below 30% of the 35 ppm spec per year. The part 
> numbers of the LM399 are not the original ones but a own manufacturer 
> specific.
> So they do a pre-ageing. Of course powered up. (perhaps intermittend 
> like on resistors?)

Interesting. And pre-selected no doubt. I wonder though if heat soaking 
at say 40C is almost, if not quite identical to powered up pre-ageing? 
If soldered to a board then its a bit different, as in normal operation 
the device is going to be warmer than the PCB. For the lower power 
band-gap type SMT references though I doubt that having it powered up is 
going to make any difference (assuming there aren't any hot components 
nearby stressing the PCB). I guess there's plenty of research out there 
on this subject.

>> I just came across another part which looks very interesting given 
>> its low cost - the automotive qualified REF5050-Q1. Although its only 
>> spec'd as 3ppm/C typical, 8ppm/C max, that's using the box method 
>> over -40 to +125C. The 'typical' chart however, figure 4, page 5 
>> shows the gradients to be very flat between 25 and 50. Its typical of 
>> course, so real parts may be very different aka Vishay foil 
>> resistors. The 0 to 85C histogram, fig 1 on page 5, do show the 
>> majority of parts being in the range .75ppm/C to 1.75ppm/C which is 
>> pretty good, and with luck, in the 25 to 50C range may well be much 
>> better so a crude heating arrangement may be worthwhile (made easier 
>> by the 5050's temperature output!)
>> I can't reconcile fig 4 with the histograms though; from the chart I 
>> reckon the 0-85 typical is approx 65ppm/85C = .76ppm/C and for -40 to 
>> 125C is approx 310ppm/165C = 1.88ppm/C. Figs 1 and 2 though show 
>> modal values of 1.25 and 2.25/2.5ppm/C. Am I doing something wrong or 
>> are these specs inconsistent?
>> Even more surprising is the headline feature on page 1:
>>  – 5 ppm/1000 hr (typ) after 1000 hours"
>> Unfortunately that seems to be an error as the 'typical' spec on page 
>> 4 is:
>> 90ppm (0-1000 hours)
>> 10ppm (1000 to 2000 hours).
>> The chart (fig 23, page 8) showing 1000 to 2000 hour drift of 96 
>> parts show the worst case being +25ppm, with the bulk ending approx 
>> between 0 and 15ppm. I wonder if they carry on improving after 2kHrs?
>> That's definitely not the SQRT(1kHr) characteristic and is very 
>> different from the standard REF5050 which quotes 100ppm (1st 
>> 100hours), 50ppm (1000 to 2000 hours).
>> If you are in a position to pre-age them for 1000 hours that 10ppm 
>> spec is almost as good as the LM399 and best of all, TI quote a price 
>> of $1.60 @ 1k parts, compared to $4.65 for LM399s @ 1k from Linear. 
>> One off prices are rather more at $4.15 from Digikey (part no 
>> REF5050AQDRQ1) but again is still a lot cheaper than an LM399 at 
>> $9.92. At $1.60 and .8mA supply current, using 4, 8 or even dozens is 
>> a realistic proposition to exploit statistical improvements and noise 
>> reduction.
>> Noise is a bit high at 15uVpp. They're also trimmable. Shame there 
>> isn't an hermetic part though.
>> Anybody tried these or spotted the gotchas? Alternatively has anyone 
>> here evaluated the hermetic LTC6655 for long term drift?
>> Tony H
> Do you really consider unheated references for a 100 ppm lifetime spec?

I'm not confident that 100ppm is realistic for < $10 after your 
responses, though I had thought that the LM399 might just achieve it. I 
didn't say I would use it unheated - if the drift over 0-40C is too 
great, despite selection, I could locate a heater resistor (or two) next 
to the reference. The reference has a temperature output and I have an 
ADC input (and enough power, hopefully not more than a few hundred 
millwatts - not done any experiments though) available so it should be 
fairly trivial to keep it within say +/-5C or better. Selecting parts 
better than 3ppm/C gives an error budget of ±15ppm or less. The real 
killer is long term drift and this part is the first I've seen spec'd at 
<= 10ppm (albeit 2nd 1000 hours ageing) apart from expensive metal can 

Also at $1.6 I could afford to do some pre-aging and selection which I 
couldn't with those $5+ devices including the LM399 and hermetic LTC6655.

> And surface mount SO-8 devices which have a factor 3 worse data on 
> hysteresis (and probably humidity) than the DIP-8 package?

Hysteresis is spec'd at 10ppm first cycle, 5ppm 2nd cycle, so not a big 
problem. They are typical figures of course so actual budget will need 
to be somewhat bigger. Since I have a microcontroller which could be 
measuring it's temperature continuously, there is the possibility to 
correct some of the hysteresis and even the temperature drift if it can 
be characterised well enough.

Humidity sensitivity is almost never published, so this is something to 
think about. Do you have any suggestions as to how much to budget for an 
SO-8 part?

Did you have a particular reference in mind in a DIP-8 package?

> I do not know where they get their "typical" T.C. curves with flat 
> area near 25 degrees. Usually they are only measuring 3 points of the 
> references. (-40, 25, +125 deg C or whatever the temperature range of 
> the device is).

They may well only measure 3 points of each device or sample batches for 
QC in production, but the device will almost certainly have been very 
thoroughly characterised when it was developed, including that 
double-humped temperature/voltage change curve in the datasheet. How 
well the bulk production parts still match that requires experimentation 
but I think its reasonable to believe they will exhibit the same curve 

> From theory any bandgap device (REF5050, LTC6655) has to have worse 
> ageing than a buried zener (LM399, AD586). But also the manufacturer 
> specific packaging has its influence. The specs in the datasheets get 
> better the later the datasheets are created for several reasons: The 
> instruments nowadays are with less tolerances. The silicon for the 
> dies is more pure than the elder qualified devices. And of course then 
> marketing demands that every generation has to be better than the 
> previous one. But this has nothing to do with physics and the actual 
> behaviour of the devices.

I've seen my fair share of 'creative' or just plain sloppy datasheets, 
but I'm not sure its fair to say that datasheets are sailing ever closer 
to the wind. I have seen a number of posts alleging that some recent 
op-amp datasheets from TI are very poor and a sad reflection of those 
from the Nat-Semi era.

It does seem that there hasn't been much progress in precision analogue 
components that push the boundaries, or even negative progress as good 
components are obsoleted - at least from the mass market manufacturers. 
I find it hard to believe that there's still nothing to touch the 
LTZ1000 after all these decades - or at least a part that approaches its 
performance but is cheaper to buy and use. No doubt due to more and more 
being done in the digital domain.

> I have my first sample of LTC6655AILS8 measured for T.C. this week. 
> But for my needs it has a too large hysteresis.
> With best regards
> Andreas
   Tony H

More information about the volt-nuts mailing list