T.C. + Hysteresis measurements on brand new LT1027DCLS8-5 voltage reference

[Gyro]

Ah, that might be a problem, the only one I have is nestling comfortably in my SVR-T. 

[branadic]

Ah, that might be a problem, the only one I have is nestling comfortably in my SVR-T. 

Well, AD586LQ is available at digikey for 13.24GBP, so if you want to invest this money in this photo it's up to you:

http://www.digikey.co.uk/product-search/en?keywords=AD586LQ

I don't have an AD586 or AD587 that I want to immolate.

[Andreas]

So here we are. This is a very first impression of the LT1027 die.
It's pretty hugh for the former microscope with x5, so I had to use another one with x3.

Hello branadic,

Thanks for the photo.
The large die might be an explanation why it took so long to package the LT1027 within the LS8 package.
(a LT1236 might be much cheaper in production).

For the photo a JQ-version of the AD58x should do the job.
But I am shure that I can find a sample within my AD586LQ which has a too large T.C. for my purposes.
(Should have thought about that before sending the LT1027#4).

With best regards

Andreas

[branadic]

But I am shure that I can find a sample within my AD586LQ which has a too large T.C. for my purposes.
(Should have thought about that before sending the LT1027#4).

You can send it every time, if you want to. To be honest, I don't have an idea how to crack them open without damaging the die. Maybe a short stroke with a hammer will crack the glas sealing?

[Gyro]

I cracked open many a CERDIP package as a kid (lots of unmarked IC bargain packs). If you find a corner where the frit seal doesn't extend right into the corner (or scrape some of it away) then inserting and twisting a flat blade screwdriver will cleanly separate the two halves without damaging the die. This used to work 9 times out of 10 anyway. Sometimes some of the legs would come away with the top half, but more often than not the entire leadframe would remain attached to the bottom half with bond wires intact.

[Andreas]

there are 2 recommendations here:

http://semitracks.com/reference-material/failure-and-yield-analysis/failure-analysis-package-level/delid-and-decap.php

or this way:
https://www.eevblog.com/forum/projects/experiment-buying-obscure-obsolete-ic-via-chip-broker/msg864675/#msg864675

with best regards

Andreas

[chickenHeadKnob]

I will do some further pictures within the next days. Also a comparision of LT1236 und LT1027 in matter of die size.

It would be interesting to see what an AD587 (or 586) looks like in comparison.

No problem, if you want to send me an exemplar preferable already opened I can take some pictures of it

 thanks to the both of you. I would like a pad/pin numbering on the photos for numpties such as myself. That is all the people need paint by numbers to follow what is going on.

[branadic]

So here is a direct comparison of LT1236LS8 (left) and LT1027LS8 (right), with pin 8 showing to the left (NC @ LT1236LS8 and Vin @ LT1027LS8).

[TiN]

Great shots, thank you.

[Gyro]

So here is a direct comparison of LT1236LS8 (left) and LT1027LS8 (right), with pin 8 showing to the left (NC @ LT1236LS8 and Vin @ LT1027LS8).

They really had to push those bond wires didn't they! 

[branadic]

Not really, but as you can see, the die is nearly twice the sice of LT1236.
And there is still enough space inside the cavity. Maybe LM399-LS8 comes next, followed by LTZ1000-LS8 

I took also some better die photos, but as the die is pretty hugh and the good microscope has x5 it tooks two pictures to visualize the chip.

[Gyro]

Wow, impressive! You can certainly see the fuses and resistors now.

[d-smes]

Why didn't they shift the die to the left in the LT1027LS8 photo to make the wire bonds easier?  Does it have something to do with the black square at the lower left?  What is that anyway?

[TiN]

It's because that patch is unplated ceramic substrate. Die need to be soldered down, and that can be done to metal surface only, hence the gold plating under the chip. Making bond wire 10mm longer is still cheaper than design and manufacturing bit different package for new chip, so they reused same one best they could.

It would be interesting to see LM399/LTZ refresh, but that's unlikely to happen due to amounts validation work required and size of that market.

[branadic]

Die need to be soldered down, and that can be done to metal surface only, hence the gold plating under the chip.

I'm pretty sure the die is not soldered into the ceramic, but glued with silver filled adhesive. Soldering requires a backside metallization (gold). Hence backside metallization is an additional process and metallization of the die is normally aluminium metallization with gold is a special-purpose step.

The indicated patch is probably for automated pick and place of the die. Squares are easy to identify with a camera as it acts as a marker inside the package.

Making bond wire 10mm longer is still cheaper than design and manufacturing bit different package for new chip, so they reused same one best they could.

Well, you can't extend bond wires infinity. At a certain lenght they will sag.
And I'm sure they know why they have decided for LS8. LT1027LS8 uses only 5 pins, LT1236LS8 uses only 4 pins and LTC6655LS8 uses all 8 pins but 4 of them are GND (we haven't open one yet to see the inside).
So maybe they choose for LS8 5x5mm² 8pin with the idea that many reference chips can be placed inside. And I'm sure there will be a batch of voltage references in LS8 package in the future. Why not LMx99 or LTZ? They only have to make sure they can fullfil the given datasheet of the devices.
On the other hand similar packages are already available for acceleration sensors. I vouch for that the package is made by Kyocera.

[TiN]

Also fixed image for your into single:



Can I use on my site for reference die photos collection? Your shots are great, I should invest time to connect my DSLR to microscope to do similar ones.
I have 6655LS8, will open one to take some photos once tests done (in some distant future...)...

[branadic]

Also fixed image for your into single:

Well done, thank you very much. Hadn't had the time for it yet.

Can I use on my site for reference die photos collection?

Sure 

Your shots are great, I should invest time to connect my DSLR to microscope to do similar ones.

Not a big deal, the microscope with cam is a Leica solution. A trained monkey can do such pictures as well, it's pretty simple. But thanks.

[David Hess]

The LTC2400 has about factor 3 more 1/f noise than the reference.

Really?  The 1/f corner for the LT1027 is not particularly low and I thought delta-sigma converters had flat 1/f noise because of input chopping.

Hmm, unfortunately Linear Technology does not include a noise spectrum graph in the LTC2400 datasheet.

Air currents can cause excess noise at low frequencies through thermocouple effects considerably larger than low frequency noise in low noise chopped and non-chopped inputs.  Jim Williams liked to use Dixie cup enclosures to ameliorate this.

[Andreas]

Really?
Hmm, unfortunately Linear Technology does not include a noise spectrum graph in the LTC2400 datasheet.

Hello,

the noise spec in data sheet is 0.3ppm.
This value is for 5 V reference with a shorted input.
And of course the value is a RMS value.
So the 1.5uVeff turn out to be 9-10uVpp.
Which compares to the 3uVpp of the LT1027.

(I searched long time for the "error" in my cirquit having around 10uVpp
until I found the 1.5uV are a RMS value).

The LTC2400 is not very low noise due to its low power consumption.
But the biggest advantages are the automated zero and full scale adjustments on every conversion.
So the LTC2400 itself is very stable over temperature.
And in my measurements I usually have the time to integrate over a full minute to reduce the noise.

With best regards

Andreas

[branadic]

Hmm, unfortunately Linear Technology does not include a noise spectrum graph in the LTC2400 datasheet.

Well, a spectrum would for sure be very interesting. Maybe Andreas can provide one, elsewise I will do some measurements in the future.

[branadic]

Can you provide a spectrum?

[Andreas]

LTC2400 noise : Already measured, the spectrum is flat, noise density about 1µV/sqHz.

Mhm,

and how does this fit to 1.5uV RMS noise for the LTC2400?
under which conditions measured?
How is the uVpp noise under those conditions.

With best regards

Andreas

[Andreas]

Hello,

a noise measurement of LTC2400 (with 2:1 LTC1043 input divider) in mV (5V input range referred to input of ADC).

1505 measurement values (within 5 minutes). so 5Hz sampling frequency.

the amplitude over time shows mostly around 10uVpp with some events (14uVpp).

FFT shows around -72dB (mV) from 0.02 to 2.5 Hz.

With best regards

Andreas

[David Hess]

and how does this fit to 1.5uV RMS noise for the LTC2400?
under which conditions measured?
How is the uVpp noise under those conditions.

It is hard to tell from the datasheet but the 1µV/sqHz EmmanuelFaure measured looks consistent with the noise specifications in the datasheet when the sinc filter shown in the "Rejection vs Frequency at Vin" figure is considered.  I had taken the noise of these converters for granted but as the datasheet points out, this is the output noise and "The output noise includes the contribution of the internal calibration operations."

I find it fascinating to consider ways to improve the noise performance.  What is the easiest way now to implement a home grown 7 or 8 digit converter these days?

[Kleinstein]

The default speed of the LTC2400 is about 6 readings per second. So at 1 µV/ sqrt(Hz), a 6 Hz Bandwidth should give about 2.5 µV RMS noise.  So my guess is the measured 1 µV/ sqrt(Hz) value includes some extra noise sources, like from the reference (not measuring at 0) or not so good anti aliasing filtering. It is also possible that noise gets higher if not measuring at 0, due the adjustment procedure. The 1.5 µV_rms specs should correspond to about 0.6 µV/Sqrt(Hz). The conversion from RMS to peak - peak values depends on the bandwidth / frequency ratio. As the readings are essentially independent it's usually a factor of 5 to 7.

There are other integrated sigma delta ADCs with a lower noise value, but often higher INL.  Some if the high resolution (e.g. 20-24 Bit) SAR ADCs from LT could be a option, though possibly with 1/f noise. Single reading noise is high, but with some averaging / integration a line period really low noise levels are possible.

Using a dual slope ADC with residual ADC is tricky, as dielectric absorption will severely limit INL, unless an exotic integration cap is used (e.g. Teflon, vacuum). So this would usually need a kind of multi slope variation, e.g. similar to the 34401. This would be more like 20 Bits from the up / down slops and maybe 12 Bits from the residual and some loss due to overlap. It's not just using many slopes, but low loss caps are much easier at 1 nF than at 100 nF.