Author Topic: EEVblog #972 - Operating Chips Outside Their Spec  (Read 24613 times)

0 Members and 1 Guest are viewing this topic.

Offline thm_w

  • Super Contributor
  • ***
  • Posts: 6278
  • Country: ca
  • Non-expert
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #50 on: February 17, 2017, 11:50:35 pm »
4.8V is min supply voltage of AD8436, not typical supply voltage.
It has to be respected.
There some reason why Analog Device has specified such a voltage : below this voltage, working of the chip is no more garanteed.

It could have been an original spec goal but has little relevance to actual chip performance, it may have been related to first batch of chips and later batches were improved, who knows.
It should be respected, but it doesn't have to be.

You say you are in QC, did you go over the original design to ensure all parts are within manuf. specifications? Did you probe the board while in various conditions to ensure all parts stay within these ranges?
Often designs will have parts running out of spec but its not intentional, it was either overlooked or unexpected. Not saying its a good thing of course, but it happens.
Profile -> Modify profile -> Look and Layout ->  Don't show users' signatures
 

Offline oldway

  • Super Contributor
  • ***
  • !
  • Posts: 2172
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #51 on: February 18, 2017, 12:05:18 am »
I do not agree.
As long I know, there is only ONE calibration value by range.

There is no calibration value for every single battery voltage below 4.8V.

It may be tested at different frequencies and signal levels on all AC range, it would be tested also at different battery voltages below 4.8V.

Ask Analog Device a special batch of AD8436 specified with 3.6V min power supply voltage or increase the battery voltage.



I hope that you will make a serious industrial multimeter, not another Chinees crap one !

Don't operate  parts outside their specifications, that's not serious, that's a principle used in crap low cost instruments.

Quote
It could have been an original spec goal but has little relevance to actual chip performance, it may have been related to first batch of chips and later batches were improved, who knows.
It should be respected, but it doesn't have to be.
You seems to know beter the AD8436 than Analog Device themself !  :-//

Why did Analog Device specify that min power supply is 4.8V if, as you say, lower power supply voltage has little relevance to actual chip performance ?

How low can we go ? 4V, 3V, 2V, 1V ?  :palm:
As you said...who knows

It does no matter if one prototype is working correctly or not !
You can't make a serious project on the base of "who knows, perhaps it works"...Specifications have to be respected in a serious product....

« Last Edit: February 18, 2017, 12:26:12 am by oldway »
 

Offline CatalinaWOW

  • Super Contributor
  • ***
  • Posts: 5173
  • Country: us
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #52 on: February 18, 2017, 12:24:52 am »

Don't worry, most of those specs for a chip are "typical" and can vary in some cases widely, with only very broad minimum and maximum values. There are a lot of parameters that are "typical" and which are only tested on a sample basis, and those often are not guaranteed, even if they are actually an important reason for choosing one chip over another in selection. Things like noise levels, actual open loop gain, true bandwidth and such are only sampled, not tested on each lot, they might only be done on a few chips from each run of wafers.


Anyone who designs based on "typical" values without some type of backup is riding for a fall.  "Typical" values are of most use to the marketing department.

I have no problem with someone doing a "typical" based design if:

a) They fully understand how the chip works and thus how it is likely to respond in untested situations.
b) They do sufficient post purchase testing to verify that the components they will physically be using are "close enough" to typical.  That requires enough analysis of their design to know what "close enough" is and what testing is required to verify that the parts installed are there.
 

Online EEVblogTopic starter

  • Administrator
  • *****
  • Posts: 37664
  • Country: au
    • EEVblog
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #53 on: February 18, 2017, 01:21:34 am »
I do not agree.
As long I know, there is only ONE calibration value by range.

Yes, but it's tested at different frequencies and voltages on each range.

[/quote]
There is no calibration value for every single battery voltage below 4.8V.
It may be tested at different frequencies and signal levels on all AC range, it would be tested also at different battery voltages below 4.8V.
[/quote]

 :palm:
There is a 3.6V voltage regulator, it doesn't matter what the battery voltage is.

Quote
I hope that you will make a serious industrial multimeter, not another Chinees crap one !

If it bothers you then you don't have to buy it.

Quote
Don't operate  parts outside their specifications, that's not serious, that's a principle used in crap low cost instruments.

The entire point seems to go completely over your head.
In a case like this it doesn't matter, as long as the meter is 100% tested.

Quote
It could have been an original spec goal but has little relevance to actual chip performance, it may have been related to first batch of chips and later batches were improved, who knows.
It should be respected, but it doesn't have to be.
You seems to know beter the AD8436 than Analog Device themself !  :-//
[/quote]

Strange that AD themselves tell us it's not a problem.
Not that it matters anyway when you 100% test and characterise something.

Quote
Why did Analog Device specify that min power supply is 4.8V if, as you say, lower power supply voltage has little relevance to actual chip performance ?
How low can we go ? 4V, 3V, 2V, 1V ?  :palm:

Again, you completely miss the point  |O
I'm done.
 

Offline NiHaoMike

  • Super Contributor
  • ***
  • Posts: 8973
  • Country: us
  • "Don't turn it on - Take it apart!"
    • Facebook Page
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #54 on: February 18, 2017, 01:36:42 am »
The Mooshimeter does RMS calculations in software with just a 32MHz 8051. In fact, it can do that on both voltage and current at the same time and also calculate real power.
Cryptocurrency has taught me to love math and at the same time be baffled by it.

Cryptocurrency lesson 0: Altcoins and Bitcoin are not the same thing.
 

Offline Brumby

  • Supporter
  • ****
  • Posts: 12288
  • Country: au
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #55 on: February 18, 2017, 05:55:34 am »
It seems all the characterization, testing and calibration under the sun isn't enough for some people, but when I see this:


Strange that AD themselves tell us it's not a problem.


any apprehensions about the use of the chip should evaporate, IMHO.


As a result, the only question that I can see in any of this is - Why did AD put that minimum voltage specification in the datasheet?

I half expect a bureaucracy based answer more than an engineering one.
 

Offline riscy00

  • Contributor
  • Posts: 22
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #56 on: February 18, 2017, 08:30:21 am »
What the model of the red oven.

What the maximum temperature it goes to?
 

Offline Kleinstein

  • Super Contributor
  • ***
  • Posts: 14080
  • Country: de
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #57 on: February 18, 2017, 08:46:08 am »
The DS from AD already shows two curves (fig 11 and 13) that show what happens when the supply is reduced. They show the errors going up / the maximum well working amplitude is going down when the curve go towards 4.8  V or a little below.

Operating at a low voltage (like 3.6 V instead of 4.8 V) is more like is would need modified specs for the rest. The more obvious point is that the optimum amplitude level is going down. It might mean the accuracy is a little lower, but as it looks, the difference is not that large.

So the main part that is missing at low voltage is the full test and specs for this. I can somewhat understand AD that they don't want to do the low voltage tests, if most users will use it with something like a 6 V or higher supply. They may end up with a different version that is the same silicon, but tested for 3-5 V instead of 4.8 -12 V.
 

Online EEVblogTopic starter

  • Administrator
  • *****
  • Posts: 37664
  • Country: au
    • EEVblog
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #58 on: February 18, 2017, 09:19:59 am »
What the model of the red oven.
What the maximum temperature it goes to?

 

Online mikeselectricstuff

  • Super Contributor
  • ***
  • Posts: 13695
  • Country: gb
    • Mike's Electric Stuff
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #59 on: February 18, 2017, 09:38:53 am »
It seems all the characterization, testing and calibration under the sun isn't enough for some people, but when I see this:


Strange that AD themselves tell us it's not a problem.


any apprehensions about the use of the chip should evaporate, IMHO.


As a result, the only question that I can see in any of this is - Why did AD put that minimum voltage specification in the datasheet?

I half expect a bureaucracy based answer more than an engineering one.
Probably to guarantee optimum performance, and operation over the whole temperature range.  Not everyone will need either.
Youtube channel:Taking wierd stuff apart. Very apart.
Mike's Electric Stuff: High voltage, vintage electronics etc.
Day Job: Mostly LEDs
 

Online EEVblogTopic starter

  • Administrator
  • *****
  • Posts: 37664
  • Country: au
    • EEVblog
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #60 on: February 18, 2017, 09:42:23 am »
It seems all the characterization, testing and calibration under the sun isn't enough for some people, but when I see this:


Strange that AD themselves tell us it's not a problem.


any apprehensions about the use of the chip should evaporate, IMHO.


As a result, the only question that I can see in any of this is - Why did AD put that minimum voltage specification in the datasheet?

I half expect a bureaucracy based answer more than an engineering one.
Probably to guarantee optimum performance, and operation over the whole temperature range.

Something like that.
The nominal 1% error graph for input level vs supply voltage goes down to 4V, so why 4.8V? Who really knows.
 

Offline f4eru

  • Super Contributor
  • ***
  • Posts: 1086
  • Country: 00
    • Chargehanger
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #61 on: February 18, 2017, 12:49:38 pm »
The Mooshimeter does RMS calculations in software with just a 32MHz 8051. In fact, it can do that on both voltage and current at the same time and also calculate real power.
at which samplerate?
seems a quite capable device. I wouldn't give up range switching and DMM Form factor. Nor a few 100hz bandwidth

Offline NiHaoMike

  • Super Contributor
  • ***
  • Posts: 8973
  • Country: us
  • "Don't turn it on - Take it apart!"
    • Facebook Page
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #62 on: February 18, 2017, 05:50:59 pm »
The Mooshimeter does RMS calculations in software with just a 32MHz 8051. In fact, it can do that on both voltage and current at the same time and also calculate real power.
at which samplerate?
seems a quite capable device. I wouldn't give up range switching and DMM Form factor. Nor a few 100hz bandwidth
8kHz on each channel. Plenty for mains frequency use. I think one of the reasons why it's not 48kHz or above is Bluetooth bandwidth limitations for the real time graphing mode. (Even 2 8kHz 24 bit channels uncompressed is pushing the limits of BLE.)

What I would really like to see is software support for an external shunt or Hall probe. The built in shunt has a 10A full scale with uA resolution which is very impressive but not quite good enough for measuring very low power microcontrollers. The option to get sub microamp resolution or being able to measure current/power greater than 10A would be very useful.
Cryptocurrency has taught me to love math and at the same time be baffled by it.

Cryptocurrency lesson 0: Altcoins and Bitcoin are not the same thing.
 

Online tszaboo

  • Super Contributor
  • ***
  • Posts: 7309
  • Country: nl
  • Current job: ATEX product design
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #63 on: February 18, 2017, 06:55:29 pm »
Expensive? Noooo, it's not! There are plenty of us today with integrated adc 12 bits or higher. Take one that is fast enough. The cost of the Micro is already included
If we calculate with a conservative (fluke 87) 20 Khz bandwidth for the TrueRMS core, you need to take samples at around 200KHz, otherwise it is not really true rms. An ADC with those specs there is the ADS8339 which is 3.5 USD from TI. You need to buffer and amplify the signal. I will assume the AD8436 is getting a 1V signal, so the amplification is only 5. Assume 0.5% max error, you need an opamp doing less than 1mV offset, 2MHz GBW. Something like a OPA197 can do it, 60 cents. You will need to dedicate an SPI port of your micro to it. Also coding needs to be done, verifying that the RMS calculations are correct. Luckily, there is only one "root calculation" in the code. I think with the usual big company impotence, a few months should be enough to develop this.
Compare this with the price of the AD8436, which is ~3USD. You gain nothing, but more chance of error for more money, and more code. While you could just solve it with an IC. #analogisbetter
Wrong price comparison. You already need a front end and an ADC anyway.
You got many advantages: a good power meter. Flexibility. Low cost. More precision on measurements (thanks to oversampling). Simpler hardware...

Nope. Analog is  not always  better, especially when dealing with lowish frequencies.

Personally, I did this exact implementation  of an rms and power calculation, at only 10 ksps, with acuracy better than 0.05% on an EFM8BB1. A 32 cent micro. With integrated 12 bit ADC!!!!!! Used in an industrial product.

I think the actual state of digital uCs and ADC allows us to make much better cheap multimeters.
A decent multimeter multiple KHz bandwidth on the AC range. Fluke 87 as a reference has 20KHz. Your implementation has less than 1 KHz, and who knows how it will handle crest factor.

The EFM8BB1 has 0.1% maximum slope error on its ADC. I highly doubt that you magically were able to create a 0.05% accuracy front end with it, I rather think you are just ignoring all the practices of proper data acquisition. You dont end up magically with better accuracy with oversampling, you get more resolution, aka more bits of the same bad measurement.
 

Online David Hess

  • Super Contributor
  • ***
  • Posts: 16548
  • Country: us
  • DavidH
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #64 on: February 19, 2017, 06:58:58 am »
Also note that you don't need absolute add precision for your 6000 counts, you can use 12 bits, for example, the massive oversampling gives you the needed precision.

This works for increased resolution but does nothing for accuracy.  The linearity and drift will remain.

How can you garantee something if you use parts out of specifications ?

Production test.

And selection, grading, and qualification.  Manufacturers make compromises in specifications for yields and test time but if it is worth it, then you can do your own testing.

Select parts through testing for a specific characteristic like transit time in bipolar transistors, low quiescent current when operating 741s above the absolute maximum supply voltage, drift over temperature, avalanche breakdown, etc.

Grade parts to match characteristics like Vbe, Vf, Vgs, hfe, ratio of capacitance, etc.

Qualify manufacturers and lots for things like insulation resistance in film capacitors, reverse recovery in standard diodes, and leakage in base-emitter junctions.

I had to write this reply twice; EEVBlog ate it the first time.
 

Offline f4eru

  • Super Contributor
  • ***
  • Posts: 1086
  • Country: 00
    • Chargehanger
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #65 on: February 19, 2017, 10:08:31 am »
Quote
A decent multimeter multiple KHz bandwidth on the AC range. Fluke 87 as a reference has 20KHz. Your implementation has less than 1 KHz, and who knows how it will handle crest factor.
my implementation can handle harmonic content up to a few kHz.
Enough for my application, 50/60Hz with only low harmonics present in any situation

Quote
The EFM8BB1 has 0.1% maximum slope error on its ADC. I highly doubt that you magically were able to create a 0.05% accuracy front end with it, I rather think you are just ignoring all the practices of proper data acquisition. You dont end up magically with better accuracy with oversampling, you get more resolution, aka more bits of the same bad measurement.
Yep. We got 0.05% typical, which is OK for a 30 cent part, and an order of magnitude better than we needed anyway.

Also, don't forget that an RMS measurement is an average of many thousands of samples, so measurement errors often tend to even out when the measurement error affects your end result in both directions.

This is not always the case. For example we found out that a small difference in sampling time between I and U channels (approx 40us) brought a small phase error that made the error in the end calculation of the active power skyrocket due to the very high reactive power always present in the application.
This error stood out because it pulled the end result in a single direction a little bit for every sample, so the integrated end result summed all these errors instead of averaging them out.
« Last Edit: February 19, 2017, 10:36:41 am by f4eru »
 

Online tszaboo

  • Super Contributor
  • ***
  • Posts: 7309
  • Country: nl
  • Current job: ATEX product design
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #66 on: February 20, 2017, 09:25:54 am »
Quote
A decent multimeter multiple KHz bandwidth on the AC range. Fluke 87 as a reference has 20KHz. Your implementation has less than 1 KHz, and who knows how it will handle crest factor.
my implementation can handle harmonic content up to a few kHz.
Enough for my application, 50/60Hz with only low harmonics present in any situation

Quote
The EFM8BB1 has 0.1% maximum slope error on its ADC. I highly doubt that you magically were able to create a 0.05% accuracy front end with it, I rather think you are just ignoring all the practices of proper data acquisition. You dont end up magically with better accuracy with oversampling, you get more resolution, aka more bits of the same bad measurement.
Yep. We got 0.05% typical, which is OK for a 30 cent part, and an order of magnitude better than we needed anyway.

Also, don't forget that an RMS measurement is an average of many thousands of samples, so measurement errors often tend to even out when the measurement error affects your end result in both directions.

This is not always the case. For example we found out that a small difference in sampling time between I and U channels (approx 40us) brought a small phase error that made the error in the end calculation of the active power skyrocket due to the very high reactive power always present in the application.
This error stood out because it pulled the end result in a single direction a little bit for every sample, so the integrated end result summed all these errors instead of averaging them out.
Yes. For 50 Hz is enought. For a multimeter, it is not, because people want to use it to debug audio circuits for example, so 20KHz bandwidth is necessary.
I've also underlined a word in your response. Sure, chips will work better than the datasheet absolute maximum. And mulitmeters are expected to do the same. So you cannot claim 0.05 if the datasheet said 0.1. And INL errors will not integrate out, because ADC will make the same bad measurement every time the signal is there. For example, you start with 3 LSB offset error. That is 0.07% of your measurement. It also states 61dB SNR for a 10KHz signal, which is about 0.1% error again. Good luck averaging that out.
 

Offline coppice

  • Super Contributor
  • ***
  • Posts: 8605
  • Country: gb
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #67 on: February 20, 2017, 11:17:46 am »
Sure, chips will work better than the datasheet absolute maximum. And mulitmeters are expected to do the same. So you cannot claim 0.05 if the datasheet said 0.1. And INL errors will not integrate out, because ADC will make the same bad measurement every time the signal is there. For example, you start with 3 LSB offset error. That is 0.07% of your measurement. It also states 61dB SNR for a 10KHz signal, which is about 0.1% error again. Good luck averaging that out.
The offset error shows up as DC. That DC component generally drifts a bit with temperature.  Any reasonable implementation of AC RMS calculation will run a continuous noise shaped DC estimator, and subtract this DC from the incoming samples before they are squared. That's your offset errors killed. The AC offset due to AWGN is not easy to completely eliminate, because it can't be continuously estimated well, and it varies with temperature, and from sample to sample of chip. It is, however easy to mitigate to some extent. Find the noise accumulated by the RMS calculation on a number of devices, at room temperature, and find their average. Now put the square of this average into the code on every device, such that it is subtracted from the accumulated squares, just before calculating the square root. You have now removed a compromise average chunk of the noise offset. Now your results should look pretty good, at least at room temperature. However, the temperature coefficient of the reference, and other parts of the MCU, are unlikely to hold the accuracy to 0.05% over a wide temperature range.
« Last Edit: February 20, 2017, 11:19:27 am by coppice »
 

Offline f4eru

  • Super Contributor
  • ***
  • Posts: 1086
  • Country: 00
    • Chargehanger
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #68 on: February 20, 2017, 07:56:20 pm »
Quote
So you cannot claim 0.05 if the datasheet said 0.1. And INL errors will not integrate out, because ADC will make the same bad measurement every time the signal is there.

Of course you can do better than the datasheet claims ! it's even the whole subject of this topic ! "Operating Chips Outside Their Spec"


And there are many ways to do it !
Doing it properly involves a risk evaluation.

One method how I would do it in this case, if the manufacturer gives no indication:
1) get a few chips with different date codes from different suppliers, samples, etc...
2) characterize them to see the lot variation on the parameter of the Datasheet. You only need to measure the parameter you want to be outside spec, in our case the INL
3) characterize the temperature variation (or voltage if variable, or any other variable external parameter for your application)
4) Calculate or evaluate which is the limit of the chip INL which will make you fail your end calculation. Let's say you could demonstrate that 0.08% INL on the chip gives 0.05% accuracy in your application, due to the sample spread when measuring AC RMS. So our limit is set to 0.08%
5) Evaluate the spread, evaluate statistically how much percent of chips will probably fail in your application (over 0.08% at the ADC).
6) Set a strategy. In our case, we decide:
- Alternative uC with better specced adc costs 10 times the price
- we produce 1000 parts a Year.
- we measured the deviation of INL from the chips in the same batch to be under 0.01%
- typical INL is around 0.04%
- we buy a reel of  EFM8BB10F2G-A-QFN20R (1500 parts) every 1,5 Years, for about 500 Euros, we could resell a roll of non useable 1499 parts for 200 Euros, but administrative effort reduces this to 100 Euros.
- we have a 2% chance to have an entire reel outside our spec
- we have an 3% chance of a marginal reel, where some parts would be outside our spec
- Product lifecycle is estimated at 8 Years -> 5 rolls needed max (count ramp up and down)
- Acuracy testing is part of our normal EOL QA testing, so characterising one chip from a new roll costs us only about 300 Euro effort, and two weeks waiting.

So we decide that a meaningful strategy is:
- we plan a reel ahead
- Every time we buy a roll, we get one chip of it, solder it manually onto a board, and pass the final test
- if ever we get a reel outside or marginal to our criteria ( 0.07% ADC, which corresponds to 0,04375% final calc), we reject the reel and resell it, re-order one, retest.

7) Talk to your boss about the added risk, and the strategy. The strategy and following decision of management have to be written down in the project minutes (very important)
8) go on and implement the strategy.


As I said already, we didn't do all that, because we didn't need better than 0.5% anyway.
« Last Edit: February 20, 2017, 08:04:56 pm by f4eru »
 
The following users thanked this post: CatalinaWOW

Offline b_force

  • Super Contributor
  • ***
  • Posts: 1381
  • Country: 00
    • One World Concepts
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #69 on: February 20, 2017, 08:04:13 pm »
I thought we were always so picky about testing things on scientific ways??
How can one sample be representative for a production of millions???

The video I saw was pretty much common sense, but also lacked a lot of information.  :--

A lot of companies are extremely conservative about certain specs, that all has to do with claims.
In some cases they even use a 3 sigma error distribution to be as safe as possible.
So first of all you use all thinkable worst case scenarios; min&max temp+max current+all I/O loaded with max current+other things that make it all bad.
In many cases this is simply being calculated from their theoretical models and simulations.
(I've worked in a company that produces IC's).
Than you compare that with a handful of practical tests, and calculate the maximum error with a 3 sigma distribution.

An other scenario is that they simple didn't test all thinkable conditions with a certain variable (the power supply voltage in this case).
Therefore they can't tell with certainty if this chip will perform according specs with lower voltages.
It's not uncommon at all that IC's or parts will perform fine outside their specs.


Offline f4eru

  • Super Contributor
  • ***
  • Posts: 1086
  • Country: 00
    • Chargehanger
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #70 on: February 20, 2017, 08:19:30 pm »
Who said one sample should be representative of millions ?

Offline CatalinaWOW

  • Super Contributor
  • ***
  • Posts: 5173
  • Country: us
Re: EEVblog #972 - Operating Chips Outside Their Spec
« Reply #71 on: February 20, 2017, 11:39:21 pm »
Quote
So you cannot claim 0.05 if the datasheet said 0.1. And INL errors will not integrate out, because ADC will make the same bad measurement every time the signal is there.

Of course you can do better than the datasheet claims ! it's even the whole subject of this topic ! "Operating Chips Outside Their Spec"


And there are many ways to do it !
Doing it properly involves a risk evaluation.

One method how I would do it in this case, if the manufacturer gives no indication:
1) get a few chips with different date codes from different suppliers, samples, etc...
2) characterize them to see the lot variation on the parameter of the Datasheet. You only need to measure the parameter you want to be outside spec, in our case the INL
3) characterize the temperature variation (or voltage if variable, or any other variable external parameter for your application)
4) Calculate or evaluate which is the limit of the chip INL which will make you fail your end calculation. Let's say you could demonstrate that 0.08% INL on the chip gives 0.05% accuracy in your application, due to the sample spread when measuring AC RMS. So our limit is set to 0.08%
5) Evaluate the spread, evaluate statistically how much percent of chips will probably fail in your application (over 0.08% at the ADC).
6) Set a strategy. In our case, we decide:
- Alternative uC with better specced adc costs 10 times the price
- we produce 1000 parts a Year.
- we measured the deviation of INL from the chips in the same batch to be under 0.01%
- typical INL is around 0.04%
- we buy a reel of  EFM8BB10F2G-A-QFN20R (1500 parts) every 1,5 Years, for about 500 Euros, we could resell a roll of non useable 1499 parts for 200 Euros, but administrative effort reduces this to 100 Euros.
- we have a 2% chance to have an entire reel outside our spec
- we have an 3% chance of a marginal reel, where some parts would be outside our spec
- Product lifecycle is estimated at 8 Years -> 5 rolls needed max (count ramp up and down)
- Acuracy testing is part of our normal EOL QA testing, so characterising one chip from a new roll costs us only about 300 Euro effort, and two weeks waiting.

So we decide that a meaningful strategy is:
- we plan a reel ahead
- Every time we buy a roll, we get one chip of it, solder it manually onto a board, and pass the final test
- if ever we get a reel outside or marginal to our criteria ( 0.07% ADC, which corresponds to 0,04375% final calc), we reject the reel and resell it, re-order one, retest.

7) Talk to your boss about the added risk, and the strategy. The strategy and following decision of management have to be written down in the project minutes (very important)
8) go on and implement the strategy.


As I said already, we didn't do all that, because we didn't need better than 0.5% anyway.

Excellent summary and a good plan.  The only thing I don't see explicitly addressed is the risk that there will be shift in production parameters which changes that 2% chance of a bad reel, or the chance of getting unlucky and getting two or more bad reels in a row.  The latter risk is easily addressed by deliberately buying the next reel very early.  At least now the cost of money is so low that things like that can be considered, but still may be a problem for your business either due to cash flow issues or opportunity cost.
 


Share me

Digg  Facebook  SlashDot  Delicious  Technorati  Twitter  Google  Yahoo
Smf