[EE] 24-bit A/D. Are We in the Twilite Zone Here?
Wagner Lipnharski email (remove spam text)
Well Dan, noise is a tricky thing.
Right now I am producing a pressure calibrator that uses the AD7713,
well, somehow extinct and replaced (supersede) by the AD7714, it is a
22.5 bits ADC, with 3 fully differential selectable inputs. The
pressure transducer supply 100mV with the full scale pressure, but if I
supply the sensor bridge with 1/3 of its required constant current, it
gives me only around 32mV at full scale. I am being reading 0.1PSI with
this 30PSI transducer, so it means around 100uV of analog signal, and
the 7713 is pretty steady and accurate.
The biggest problem I found years ago with low voltage is really noise
in voltage reference generators. At this point I found out that in this
particular point "you have what you pay for". The LM358 1.2 and 2.5V
reference has several versions, going from 91 cents up to $4.66
(30ppm/degC). The LM4040 series are pretty good too, but you need to go
for a .1%, as the LM4040AIZ-2.5-ND (Digikey) for $4.10 per (TO-92)
unit. The most important part of the tech specs of those vref's is the
"noise" part. Some of them just obscure it, other just misinform, but
some state it. Some of them have milli-volts of noise (white noise,
pink, whatever), it is an intrinsic silicon noise, can't be removed,
can't be filtered, there is NO capacitor can avoid it, stabilize or
level it. You just need to check the manufacturer document (specs
sheet) for the one you are using. The above information is from
Digikey, you can get better products around.
Years ago I made a nice portable instrument, one of the functions were
to generate current and voltage, in micro-volts range. I suffered like
hell trying to locate why the output voltage was so noisy. The output
was programmable from 0 to 25V in 10 micro-volts steps, so you could
specify 24.999,99 but the high precise HP unit I was using to measure it
always showed last two digits floating and jumping. I changed
everything, from the solderless protoboard (first mistake), op-amps,
capacitors, resistors to 20ppm 0.1%, and so on. Using a signal booster
I could see the vref noise. I was using a 1.25Vref, noise was in the
60uV range (according to the manufacturer), but the DAC was multiplying
by almost 20 times to get the 25V output, so the noise went up to almost
1.2mV!!! sometimes higher.
So I learned that there are two very, very important things in all
electronics... the bases... Clocks for digital and Vref's for analog.
They can cost up to 1/20 of the total equipment, now I recognize and pay
it with pleasure, they ARE what the product will offer as quality and
accuracy anyway. Doesn't make sense to use a 91 cents Vref in a product
that cost $1500 to the customer, and that cheap Vref will define the
product accuracy and quality... you know what I mean.
I never thought to use a thermal oven for a crystal oscillator in life,
but I did once. It runs under 45¡C, needs 48 hours to stabilize before
calibration, and needs 4 calibrations, once each 6 hours after the
initial warm-up. Can't use a regular metallic screwdriver to touch the
cap inside the crystal chamber since it can change the internal temp...
and of course frequency, isn't that funny? The plastic screwdriver
should be kept in another 45¡C chamber... scary. But when you go for
critical and high accurate things, there is no other way.
Calculations: 1.25Vref with 60uV of noise, means:
625mV = 1 bits
312mV = 2 bits
156mV = 3 bits
78mV = 4 bits
39mV = 5 bits
19mV = 6 bits
9mV = 7 bits
4mV = 8 bits
2mV = 9 bits
1mV = 10 bits
610uV = 11 bits
305uV = 12 bits
152uV = 13 bits
76uV = 14 bits
38uV = 15 bits
So, the noise kills the unit at 14 bits, can't use it with a 16 bits ADC
Now, other thing. accuracy (percentage of error).
The LM4040 unit described above, has 0.1% or error. It means that in
1.25000V vref, it can be in real 1.249000 or 1.251000, now, if you look
at the above table, find the 1/1000 part, you will see the 1mV at 10
bits, so all above 10 bits will be not valid.
Now, look for ppm, a 30ppm/degC, means the unit will change its value in
30 parts per million per each celsius degree the ambient temperature
changes from the based 20¡C. Then, suppose the room is at 25¡C, it will
be 150 parts per million, 150/1000000 or 0.00015, in other terms it is
0.015%, pretty high for only 5¡C in change, it is 0.003% per ¡C. The
original 1.250000 V will change to 1.249250 V at zero ¡C, add a possible
.1% of negative error, will go to 1.248V, so it is already 2mV, drop
your accuracy to below 9 bits... pretty scary, huh?
We are not talking about a cheap 91 cents Vref, but about the 30ppm (one
of the best at the market) LM4040, with the best possible available at
Digikey with 0.1% of error, the most expensive one, around $5 per unit,
and that thing can't really goes up to 12 bits at 0¡C???? Yes!
What happens here is that you need to deal with all the possibilities.
First, you need to compensate and should trust in calibration. If the
unit is actually with 0.1% of error, the calibration will fix it, IF the
unit keep the error permanently. Second, temperature offset and
deviations should also compensated by software, for that you need to use
a temperature sensor and calibrate the unit for different temperature
ranges, so the software would know where to look and how to do.
Going to your problem.
2mV noise is pretty high. How do you get it? How do you measure it? at
the ADC output? or with an oscilloscope?. What is your signal source?
what impedance it offers? To work with very low signals, it is *VERY*
important to match impedances, if not, bounced signals can kill the
original and clean one. Sometimes a capacitor creates more problems
than solve. A capacitor is not a God's creation. It leaks, it farts, it
dirts all over the place in very low voltage.
At 24 bits, nanoVolts range, a single electron leaking at the cap plates
can creates a bouncing signal at the ADC input.
Here again, expensive and good polypropylene caps makes a hell of a
difference from cheap ones.
What happens if you short circuit the ADC input, still having the noise?
The trade-in here is to avoid at any cost high gain amplifiers, try to
deal with the signal as pure as it is, avoid intermediate circuits,
resistors, dividers, filters, whatsoever that can transform your life in
a live hell. Signal treatment is a long and extensive learning at any
electronic university, and it is not any teacher of this subject that
can build a nice front-end amplifier to work with a 24 bits A/D.
Answering your subject question; Yes, at 24 bits A/D, we really ARE at
the twilight zone, doomed, alone, in total fear, no support, no
experience around, no books, no fancy "do it by yourself multicolored
animated gif playing at your screen"... just plain despair!
PS: Just try and turn off the PC and that fluorescent lamp over your
bench, and see if the 2mV disappear... :)
Dan Michaels wrote:
See also: www.piclist.com/techref/io/atod.htm?key=a%2Fd
You must be a member of the
piclist mailing list
(not only a www.piclist.com member) to post to the