Are you really fine with IC voltage regulators ?

Long time members of IEEE's Solid State Circuits society, had the option to purchase the first 55 years of ISSCC digests (plus the first 40 years of JSSC journal issues) on DVD-ROM, back when DVD-ROM was cheap and 1 megabit internet access was unobtanium. I am one such person and own them; they are currently, uh, somewhere in my downstairs rabbit warren.
 
I am looking at various options to implement a voltage reference for ADC's and DAC's.
One of the options is to use an LM431/TL431 + filter + buffer.
For this I have searched for '431's with low noise, reducing the requirements for the filter.
In the table in post #497 link and indeed in the datasheet (curve) from onsemi, the noise is specified as 48 nV/sqrt(Hz).
Has anyone here or somewhere else been able to confirm this number?

I have just made some measurements, where I got a noise of around 123 nV/sqrt(Hz).
This was with the TL431BCLPRMG (TO-92) from onsemi.
The noise was calculated from a total noise of around 18.2 uV over a 22 kHz BW.
Interestingly, the noise spectrum was flat, with no 1/f noise visible down to 10 Hz.

With the LM431SBCM3X (SOT23) the total noise was slightly lower at 16.5 uV, equal to 112 nV/sqrt(Hz).
The 1/f corner was around 100 Hz.
For the LM431 the noise is not specified. Only mentioned under features as "Low Output Noise Voltage".
One of my goals was to find out if the TL431 and LM431 from onsemi were comparable, since no noise data is available in the datasheet of the LM431.

I have seen mentioned before that the onsemi versions should have lower noise than comparable '431 devices, but the values I measured seem to be similar to the approximately 125 nV/sqrt(Hz) specified by TI.
 
  • Like
Reactions: dreamth
Here's the way I tend to look at it: current sources are by nature high impedance. High impedance means higher thermal noise as compared to low impedance. However, Cset can filter out most of that noise, so the end result can be good. Doesn't mean the current source by itself is inherently low noise. Means it needs a good filter. Is that wrong?
 
In small signal incremental analysis, the output impedance of the current source appears in parallel with "Rset", the resistor which the current source drives. Since Rset is at least two orders of magnitude smaller than the current source output impedance, the parallel combination of these two impedances is therefore Rset. It is Rset which sets the Johnson-Nyquist noise magnitude at the reference input. The purpose of the bypass capacitor is to shunt Rset's noise to ground, thus keeping it away from VREF. The optimum value of Cset is infinity farads but in some applications this may be impractical.
 
Sure. IIUC the current source can be noisy but its not necessarily a problem given Cset and Rset form the corner frequency of the filter (this assuming Rset is much smaller than the effective current source equivalent resistance). Don't see how MarcelVDG misspoke? Did he?
 
Last edited:
Point taken, and always understood. Perhaps the point I was actually trying to make and didn't express well enough was missed. The noise of the current source is of little matter. The current source may in fact be noisy when considered alone. However if so stated as a matter of fact, is there something wrong with making the observation that it happens to be noisy? That is the question that prompted my initial reply.
 
SPX2431 may have slightly lower noise but not a big difference to TL431.
Thanks. I have just ordered some SPX2431 to test.

After doing so, I found a discussion here, which seems to verify that they have slightly lower noise than the TL431, but not much. And still way above the datasheet level.
96 nV/sqrt(Hz) was measured in that case.

And here the noise of the TL431 from onsemi was measured at 127 nV/sqrt(Hz). This is close to the 123 nV/sqrt(Hz) that I measured.

I have a suspicion that what onsemi specify (48 nV/sqrt(Hz)) and Sipex (now MaxLinear) specify (38 nV/sqrt(Hz)) is the noise per volt of reference!
123 nV (my measurement) divided by 2.5 is 49.2 nV and 96 nV (bcarso's measurement) divided by 2.5 is 38.4 nV. Very close to 48 nV and 38 nV! Probably too close to be a coincident. But perhaps they forgot to tell that in the datasheet?
 
  • Like
Reactions: chris719
Last edited:
I suppose voltage reference is not for some instrumentation DAC and some small and slow temperature related drift is not a problem. If so, why not to consider Walt Jung’s LED based voltage references. Noise density for a basic 2.5 V GLED reference is from 2 to 4 nV/sqrt(Hz).
Your assumption is not quite correct. I do intend to use it for instrumentation purposes 🙂
So I need low noise as well as a decent precision and stability. But the precision and stability does not have to be at the LTC6655 level. And of course I would prefer a significantly lower cost.
 
@JensH, I've been using Walt Jung references for a long time and I think they're worth a lot more than that one ZTX and the green Lite-On LED.
An ordinary zener diode can be used instead of an LED if you want to achieve precision, but then the noise is higher .
If you are interested in the exact temperature drift of the GLED431, I can measure it for you this week, I use them in all regulators for 5V in my DAC.
 
The best LED, noise wise, that I have seen is the hlmp-6000.
< https://www.digikey.de/de/products/...diskret/105?s=N4IgTCBcDaIBYBsC2AHAbABiyAugXyA >

Originally HP, then Avago , then Broadcom
Optically it's quite a dimm bulb.
Someone else here on diyA gave me the tip.
0 dB is 1 nV/rt Hz, bias is 1 or 2 KOhm NiCr wire from 14V NiMH cells.
The trace with the 47R shows that the bias supply does not add any significant noise.
 

Attachments

  • 24354944411_007c4fe5db_o.png
    24354944411_007c4fe5db_o.png
    38.1 KB · Views: 147
Last edited:
The purpose of the bypass capacitor is to shunt Rset's noise to ground, thus keeping it away from VREF. The optimum value of Cset is infinity farads but in some applications this may be impractical.
No. The optimum value of Cset is 100 uF tantalum. More brings no benefit.
Measured in a 5Vout system. That is already more than the data sheet says.
And Cout > 4u7 brings only a bigger noise peak at 10 or 100 KHz, depending
on current.

Gerhard
 
I find post #557 both illogical and practical. I would have regarded it as both logical and practical if you had written something like: 'Mark is correct in theory, but for frequencies of 10 Hz and higher, 100 uF is close enough to infinity' followed by the explanation of what you measured with what type of capacitor.
 
  • Like
Reactions: SonocusEric