Quick sanity check needed on an Engineer's Thumb schem

[suryabeep]

Hi everyone,
I drew up a schematic for the Thumb with all the mods and stuff, can someone have a quick look if there are any blatant errors before I send it off to the fab?

[Rob Strand]

if there are any blatant errors before I send it off to the fab

Blatent would be no resistor to 4V5 on pin 3 of IC2A.

[samhay]

I assume the threshold pot is supposed to set the bias for IC2A. At the moment it is blatantly not right.

[rankot]

There's a schematic with paralleled LM13700 which reduces noise, you can find it on original thread.

[suryabeep]

Rob and Samhay - Thank you! I'll fix them on this version
Rankot - this one? https://drive.google.com/file/d/0B1wLPwWafEnXTkRKMl93UVZkVms/view Seems pretty cool, I'll make this v2.0 lol. I'll draw it up in eagle (and probably need another sanity check  ) Just to confirm, U1 is the regulator, U2 is a quad-opamp, U3 is the LM13700, and Q1 is the BS170?

[rankot]

I'm not sure it's the most recent version - that is certainly the one combined with Fencepost noisegate. You can use that one. I don't have Eagle files, but if you decide to make your PCB using PCB Creator, I can send you that file.

[rankot]

U1 is TL074, U2 is NE5534, U3 LM13700 and Q1 is BC327 (for LED compression indicator, but it's not working as I expect it to do).

[suryabeep]

Rankot- Is this the latest version? http://i67.tinypic.com/2moeek8.jpg
If so, yeesh that's huge. I think I'll build the small one for now 
Thanks for the clarification about the IC numbers. What's the U1 in the top that's only got V+ and V- connections? I see U1.2, U1.3, and U1.4 but no U1.1 - Is the 'U1' really the U1.1? I'm not sure I understand this scheme fully 

As for the original scheme I posted, the only fix needed was the IC2A 4V5 connection right?

[rankot]

U1 with only two (+&-) connections is just schematic representation of TL07x Vcc/Vdd.

This is complete schematic with paralleled LM13700:
https://www.dropbox.com/s/fikr12lidgeg69f/ETparallel.pdf?dl=0

TL072, NE5534 and LM13700.

[DrAlx]

Does using 2 OTAs that take half the bias current each really lower the noise level compared to having just one OTA taking all the bias current?
Has anyone measured the improvement with a scope?

I don't know what the dominant source of noise is from an OTA.  I assume its shot noise from Ibias.  If that's the case then I don't see how splitting Ibias between 2 OTAs and then summing the two output currents helps.

i.e.

With a single OTA with current Ibias you get mean square current noise of

  <Inoise^2> = q* Ibias     (per Hz).

With two OTAs, each taking half Ibias, then <Inoise^2> for each OTA is half the single OTA case.  When you sum the currents back at the OTA outputs, you get the same output current as before and the same mean square current noise as before.

[Rob Strand]

Does using 2 OTAs that take half the bias current each really lower the noise level compared to having just one OTA taking all the bias current?
Has anyone measured the improvement with a scope?

Back in 2000 1996 or so I used it on NE570/NE571/NE572s. I found those chips have quite a bit of feedthrough when the gain control is near maximum so I backed-off the signal level and used two in parallel to counteract the poorer S/N.   Band-limited noise measurements confirmed the noise is reduced.   The idea is the noise from each VCA is uncorrellated so by using two the noise increases by 3dB but the signal increases by 6dB so you gain 3dB S/N.

I also found another way of reducing noise in the NE570/NE571/NE572s.   Some connections caused a stinking load of noise to get through.  It was something to do with filtering Vref with an RC filter and carefully adjusting the filter resistor value.  [Forgot to add this wasn't just changing the RC filter's cutoff, there was a secondary noise-gain mechanism going on which was related to the value of the resistor.  The Vref tap is required when using an external opamp.]

I was convinced it was the best NE570/NE571/NE572 compressor out there.

[DrAlx]

The idea is the noise from each VCA is uncorrellated so by using two the noise increases by 3dB but the signal increases by 6dB so you gain 3dB S/N.

I get how that VCA example works.

With one VCA:
Signal Voltage = V
<Vnoise^2>  = N
SNR = (V*V) / N     (squaring voltages to get power ratio).


With two such VCAs:
Total signal voltage = 2V
Total <Vnoise^2> = 2N
SNR = (2V*2V) / (2N)   = 2*(V*V) / N     (hence 3dB improvement).

That's not the case if the noise is proportional to the bias current (e.g. shot noise).

With one OTA:

Signal current = I
<Inoise^2> = q I   
SNR = (I*I) / <Inoise^2>  =  I / q.    (squaring currents to get power ratio)

When you have 2 of these OTAs and they are *not* sharing the bias current

Total signal current = 2I
Total <Inoise^2> = 2qI
SNR = ( 2I * 2I ) / ( 2qI ) = (2I)/ q.

So you get 3dB improvement as in the VCA example, but this is only because you actually have double the current that you had with a single OTA.   

If you split Ibias, you don't get that improvement.
Instead you get

Total signal current = 2* (I/2) = I
Total <Inoise^2> = 2 * (q * I/2) = q I
SNR = ( I * I ) / ( q I ) = I/q.


Now I don't know how similar the VCA you looked at is to an OTA. 
That's why I was asking if anyone had measured with this OTA example.
I assumed that OTA noise power is proportional to the bias current (like shot noise) but it may not be.

EDIT:  A simpler way of saying this is that I am assuming that the OTA SNR is proportional to Ibias.
That would make sense to me because the noise in this circuit is greatest when the OTA is being driven with
small Ibias (i.e. when the compressor is trying to provide the largest gain).

2nd EDIT:  Actually its more complicated than I thought. I've been a bit loose with the terminology and treating Ibias almost like its the signal term because it multiplies the signal voltage at the OTA input.  The OTA will actually produce more noise when you run it at larger Ibias, but in this circuit it's really acting like a (noisy) variable resistor.  So do you get a less noisy resistor with two of these running at half Ibias?   Need to have a think about this.

[Rob Strand]

Now I don't know how similar the VCA you looked at is to an OTA.
That's why I was asking if anyone had measured with this OTA example.
I assumed that OTA noise power is proportional to the bias current (like shot noise) but it may not be.

There's also some component due to the Rbb' of the OTA transistors.   That component is always uncorrelated.

Most "gain-cells" and OTAs work on the same idea.  They are some sort of differential amplifier where the tail current is controlled and varies the transconductance.  So yes, the NE57x chips are representative of the OTAs. 

When you have 2 of these OTAs and they are *not* sharing the bias current

Because of this, connecting two OTAs might introduce some correllation.  If the Rbb' part isn't dominant and correllation occurs then the S/N improvement might not be the full 3dB.

Equation (36) of this paper gives the noise voltage proportional to the control current:
http://ajoliveira.com/ajoliveira/gen/pdf/preprints/paris88.pdf

[Rob Strand]

Actually its more complicated than I thought. I've been a bit loose with the terminology and treating Ibias almost like its the signal term because it multiplies the signal voltage at the OTA input.  The OTA will actually produce more noise when you run it at larger Ibias, but in this circuit it's really acting like a (noisy) variable resistor.  So do you get a less noisy resistor with two of these running at half Ibias?   Need to have a think about this.

Yes, it's not *that* simple. If you just think about non-correlated sources the idea is easy. When you start to come up with a real model with all the interractions (which cause correlations) is gets tricky.  I remember going through these exact thoughts when I first came up with the idea.  The easy answer is it will be a  less than 3dB.

I worked on an ultrasound machine which uses leading-edge mixers (which use transconductance control) to deal with the insanely small signals.  The noise from the mixers is very close to theoretical - a tiny Rbb' component.  I seemed to remember little was to be gained by a parallel connection.  I actually tried it and got no improvement.  Unfortunately that was before I found other noise sources which could have masked the improvement.  I never got back to trying the parallel connection after all the other noise sources were eliminated (some were numerical not electronic.)

[DrAlx]

Equation (36) of this paper gives the noise voltage proportional to the control current:
http://ajoliveira.com/ajoliveira/gen/pdf/preprints/paris88.pdf

Thanks Rob.  Using two OTAs makes sense to me now.

[PRR]

> Rbb' part isn't dominant

We usually drive '3080/'13700 with a couple dividers, 220r. Rbb' is surely <100r, perhaps <50r. So Rbb' is not real dominant.

If not for that (ALL other factors): yes, doubling the Area of the input devices, at the same total current, leads toward -3dB hiss Voltage.

[Rob Strand]

We usually drive '3080/'13700 with a couple dividers, 220r. Rbb' is surely <100r, perhaps <50r. So Rbb' is not real dominant.

Good point.     I suspect the on-chip Rbb' are quite high.  The LM394's final Rbb' was 40 ohms and that has quite a number of parallel transistors.  OTA's like those from THAT are likely to have nice low Rbb's.   The NE57x's could very well be high.

If not for that (ALL other factors): yes, doubling the Area of the input devices, at the same total current, leads toward -3dB hiss Voltage

The complication is there's cases were paralleling doesn't always decrease noise.  For example if we have a transistor stage and replace the transistor with two in parallel transistors (assuming perfect matching and current halving).  The Rbb' of the final transistor decreases but the shot noise stays the same. 

It's like,
Ic = collector current for single transistor case.
IN1^2 = 2q (Ic/2) B        ; shot noise Q1
IN2^2 = 2q (Ic/2) B        ; shot noise Q2
Itotal^2 = IN1^2 +IN2^2  = 2q I c B
which is the same shot noise as a single transistor.
So we haven't gained anything from a shot noise point of view.

If we get two identical stages with noiseless buffers at the input, then combined the outputs vo = (1/2)*(v1+v2) then I believe we improve S/N even for Rbb' = 0.     For this case it's easier to use the equivalent input noise voltage model.   For Rbb'=0, the input noise voltage *decreases* with increasing current.  If we half the bias current in each stage, and adjust the voltage gain to be the same as the higher current stage,  then the signal to noise ends-up being the same as the single stage.

I suspect putting two ideal (Rbb'=0)  OTAs in parallel, each operating at 1/2 IABC, ends up giving the same final noise like the parallel transistor case.   The only gain is the parallel Rbb's.    Something I need to think about more is for the OTA, with Rbb' = 0, the input noise voltage is inversely proportional to the tail current.  However the transconductance  is proportional to the tail current.  So the noise current for Rbb'=0 should be constant.

I noticed something specific about the NE57x's  The gain-cell has a stage before the transconductance stage.  The noise of the first stage  is independent of the control current.
So I might have to retract my remark that NE57x's are the same as OTA's.