Tee networks in TIA feedback loops

[Nemo]

I have a high capacitance photodiode (650pF with as much bias as I can
put on it) which I want to measure at high gain and 100kHz bandwidth or
more. It works well enough to meet spec, but I'm trying to make it
better simply to learn more about analog design. Simply using a TIA to
amplify it by about 1M - I'm trying to get as much gain in this first
atge as possible to reduce Johnson noise in the feedback network - I get
minor oscillations at around 100 - 150kHz unless I increase the feedback
C to a point where my bandwidth is adversely affected. I was wondering
about using a T network to get round this, but my experiments didn't
seem to improve the instability, and I wondered if there was something
obvious I was missing.

I should say first that I rapidly found that there was a lot of DC drift
from using a T, so if doing this on a PCB I would use one where the
elements are on the same chip and had tight tempco matching (Vishay do
some with 2ppm matching, in a SOT23 package. Expensive, but price is not
a concern here.)

The circuit I tried was along these lines:

_____1 - 3pF_____
| |
| ___100k_______|
| |
i --> --|__|\ |
| \______10k---1k___
0V__| / |
|/ 0V

I reasoned I would be able to use a higher C, maybe 10 - 20pF in
parallel with the 100k to get stability, but maintain the same
bandwidth. But I found that increasing the feedback C to the point where
the instability is reduced, just decreases the bandwidth dramatically,
as if the C is also being multiplied by a factor of 11.

The amplifier is a compound one along the lines Phil Hobbs suggested a
few weeks ago: JFET front end feeding low voltage noise op amp, with
auto nulling circuit servo-ing the input to 0V. I've tried several op
amps of various GBW's and all display this slight ripple in the output.

I concluded that T's are good if you want an unfeasibly high feedback
resistance, but have no bandwidth advantage. But I hope I'm missing a
trick here and you guys will drop a hint 8)

I also wondered what is really meant by a "stable" amplifier. Students
are told "stick more C in the feedback loop to make it stable." Is this
one of those glib simplifications? I'm looking at the output at high
gain and there is always _some_ ripple present if there is significant C
on the input and you turn the gain on the 'scope up high enough. Surely
there will always be SOME degree of chase-your-tail hunting round the
feedback loop for any amplifier? Or is it possible to get a TIA which is
actually truly stable with this kind of C on its input?

Nemo

 

[Helmut Sennewald]

I have a high capacitance photodiode (650pF with as much bias as I can
put on it) which I want to measure at high gain and 100kHz bandwidth or
more. It works well enough to meet spec, but I'm trying to make it
better simply to learn more about analog design. Simply using a TIA to
amplify it by about 1M - I'm trying to get as much gain in this first
atge as possible to reduce Johnson noise in the feedback network - I get
minor oscillations at around 100 - 150kHz unless I increase the feedback
C to a point where my bandwidth is adversely affected. I was wondering
about using a T network to get round this, but my experiments didn't
seem to improve the instability, and I wondered if there was something
obvious I was missing.

I should say first that I rapidly found that there was a lot of DC drift
from using a T, so if doing this on a PCB I would use one where the
elements are on the same chip and had tight tempco matching (Vishay do
some with 2ppm matching, in a SOT23 package. Expensive, but price is not
a concern here.)

The circuit I tried was along these lines:

_____1 - 3pF_____
| |
| ___100k_______|
| |
i --> --|__|\ |
| \______10k---1k___
0V__| / |
|/ 0V

I reasoned I would be able to use a higher C, maybe 10 - 20pF in
parallel with the 100k to get stability, but maintain the same
bandwidth. But I found that increasing the feedback C to the point where
the instability is reduced, just decreases the bandwidth dramatically,
as if the C is also being multiplied by a factor of 11.

The amplifier is a compound one along the lines Phil Hobbs suggested a
few weeks ago: JFET front end feeding low voltage noise op amp, with
auto nulling circuit servo-ing the input to 0V. I've tried several op
amps of various GBW's and all display this slight ripple in the output.

I concluded that T's are good if you want an unfeasibly high feedback
resistance, but have no bandwidth advantage. But I hope I'm missing a
trick here and you guys will drop a hint 8)

I also wondered what is really meant by a "stable" amplifier. Students
are told "stick more C in the feedback loop to make it stable." Is this
one of those glib simplifications? I'm looking at the output at high
gain and there is always _some_ ripple present if there is significant C
on the input and you turn the gain on the 'scope up high enough. Surely
there will always be SOME degree of chase-your-tail hunting round the
feedback loop for any amplifier? Or is it possible to get a TIA which is
actually truly stable with this kind of C on its input?

Nemo

Hello Nemo,

The drawback of every T-network is higher noise and offset drift.
So there there are only drawbacks for yor TIA. Don't use a T-network.

Helmut

 

[Joerg]

Phil Hobbs wrote:

[...]

It's also a win to bootstrap the bootstrap, and to make the bootstrap
gain as close to 1.0000 as you can--I just finished a
bootstrapped-bootstrap design where the first stage output is from the
bootstrap, rather than from the hot end of the PD--no TIA stage
required. It'll work beautifully if we can guard out the stray
capacitance to ground.

That reminds me of a financial market comment on Limbaugh's channel:
"And then they leveraged the leverage that was already leveraged"

But yeah, bootstrapping can be great.

 

[Joerg]

Phil Hobbs wrote:

[...]

It's also a win to bootstrap the bootstrap, and to make the bootstrap
gain as close to 1.0000 as you can--I just finished a
bootstrapped-bootstrap design where the first stage output is from the
bootstrap, rather than from the hot end of the PD--no TIA stage
required. It'll work beautifully if we can guard out the stray
capacitance to ground.

That reminds me of a financial market comment on Limbaugh's channel:
"And then they leveraged the leverage that was already leveraged"

Increases the noise.

It was more like *KABLAM* in fall 2008.

 

[Joerg]

Phil Hobbs wrote:

[...]

It's also a win to bootstrap the bootstrap, and to make the bootstrap
gain as close to 1.0000 as you can--I just finished a
bootstrapped-bootstrap design where the first stage output is from the
bootstrap, rather than from the hot end of the PD--no TIA stage
required. It'll work beautifully if we can guard out the stray
capacitance to ground.

That reminds me of a financial market comment on Limbaugh's channel:
"And then they leveraged the leverage that was already leveraged"

But yeah, bootstrapping can be great.

Bootstrapping the bootstrap is the only way to get the effective
capacitance below C_DG of the BF862(s). This particular design has a
boostrap gain of about 0.995ish, and holds up well over frequency, so
that the bandwidth goes up by about 130 times compared with the plain RC.

130 is impressive.

It still has the differentiated noise, and we have to keep the
capacitance to ground at the hot end of the photodiode about 0.1 pF or
less. That ought to be quite doable with a bootstrapped island in the
ground plane (big enough that the fringing doesn't dominate) and a small
piece of copper-clad polyimide over top, also bootstrapped.

Because the bootstrap gain is so nearly 1.0, there isn't a lot of
opportunity for nonlinearity, so I'm taking the output from the (low
impedance) bootstrap, which I haven't seen done before. (Of course I
don't get out much.)

High dynamic range design is fun.

Yup. Done it for most of my career because instantaneous dynamic range
in medical ultrasound Doppler has about the same importance as
horsepower does in the world of muscle cars.

 

[Nemo]

With high-value feedback resistors, T networks are not a win. The
resistor noise dominates the amplifier noise, so there's no reason to
accept the higher offset voltage.

Whether there's a SNR hit or not depends on how you do it. If the
optimal design needs a 1G feedback resistor, and you decide to use a
100:1 voltage divider feeding a 10M feedback resistor, your input
current noise just got worse by 20 dB, which is bad.

On the other hand, feedback T networks can help SNR in some situations,
by reducing the second-stage noise contribution. This is mostly an
issue with feedback resistors smaller than 10k. The key thing is to
keep the same (i.e. optimal) feedback resistor, and use the T network to
get a bit of voltage gain from the stage, to override the second stage's
input voltage noise.

I don't have significant voltage noise in the subsequent stage, and
first stage gain is "only" 1M, so it sounds like Tee's are no help. I
found a Pease Porridge article which mentioned hey are SOMEtimes
justified, at low ratios of maybe 3:1, to help tame diode capacitance
which is why I wondered about this approach. But it doesn't seem to
quell the instability.

You're stuck with the differentiated input noise of the amplifier,
regardless, but you keep winning by parallelling N BF862s until either
your power supply melts or the gate-drain capacitance starts to dominate
the PD capacitance. (You win lower noise by sqrt(N), but lose linearly
in N once the capacitance gets too big.)

Yup I figured that from your post a few weeks ago 8)

It's also a win to bootstrap the bootstrap,

I'm afraid I don't know what you mean by this. So far I've simply been
using a BF862 source follower as per one of the Linear Tech app notes
(BTW did you know the LSK170C from Linear Systems is, on paper, a second
source for this - but in practice it seems to be perhaps a shade
noisier). I've tried running it at various currents as JFETs are meant
to be slightly quieter at high currents (can't say I've found any
difference, maybe my other noise sources are swamping the effect).

and to make the bootstrap

gain as close to 1.0000 as you can--I just finished a
bootstrapped-bootstrap design where the first stage output is from the
bootstrap, rather than from the hot end of the PD--no TIA stage
required. It'll work beautifully if we can guard out the stray
capacitance to ground.

It's a shame I live on the other side of the Atlantic or I would be
interested in that lab assistant post you mentioned, even at a cut in
pay. You do some fun stuff and whoever gets that job will have fantastic
experience.

 

[Nemo]

Are you bootstrapping out the photodiode capacitance? Bootstrap plus
cascode can improve bandwidth by 100:1 or so. Phil has some notes on
his web site, and more in his book.

Bootstrap: yes, but reading what Phil says I realise I need to improve
that. Enormously.

I have his book and the cascode buffer was the first thing I tried.
However, it's best for signals in the uA region. I'm trying to resolve
down to maybe 100pA and the shot noise of the 7uA quiescent current
through his cascode transistor has a shot noise of about 500pA over
100kHz. And I found problems with DC drift because it's sensitive to
tiny fluctuations in supply voltage. These can be ignored for a gain of
1M but my overall gain is 1G and they show up quite strongly. It's a
measure of how bad my first breadboard was that it still improved my
S/N! I've certainly learnt a lot about all manner of analog on this
journey. I tried cutting corners like in the power supply at first...

 

[Nemo]

Phil described a bootstrapped bootstrap thus:

You hang an emitter follower off the JFET source, and use that to
bootstrap the drain of the JFET. Use _big_ capacitors, because you care
about the phase shift.

For decoupling, I generally use a 10uF X7R in parallel with a 10nF
COG/NPO to catch the high freq stuff - is this what you mean by big?

Take the PD bootstrap cap from there too. If

you expect short pulses, pick the BJT polarity so that the follower
turns on harder when you get a pulse of light.

The major source of voltage error is the limited g_M of the FET--it's
about 30 mS, but that's a factor of 10 less than a BJT at the same
current. Using a current source load for the FET helps a lot, but make
sure it's a quiet one. It isn't hard to make a sub-Poissonian current
source, you just use a regular BJT with fixed base bias and a few
volts' drop across its emitter resistor. You win by running the
followers fairly hot--15 or 20 mA is often best. Be sure to bypass the
base to the negative supply, or the supply noise will come right in on
top of your signal.

Yes I've only recently grasped the decouple-to-negative trick 8)

(There's a poor-man's version of this, which is to run the FET and BJT
like a Darlington, with the FET's source load being a small resistor R_P
from base to emitter of the BJT, so that the drain current is nearly
constant at V_BE/R_P. H&H have that one, I'm pretty sure. You don't
get as high a source load impedance that way, and the shot noise of the
BJT corrupts the low noise of the FET.)

Oh, and remember to use a two-pole capacitance multiplier on the PD bias
supply, and crank up the bias as far as you can.

I'm already doing those 2 items (learnt through earlier overconfidence)

Thank you, this is all very educational. Much appreciated.

 

[Joerg]

In a high performance bootstrap, _everything_ is a big deal. The
difference between an AC gain of 0.95 and 0.995 is a factor of 10 in
bandwidth.


I agree. In this case, though, you need to have the impedance level
dropping ~100x per stage, so the earlier stages are way too wimpy for
feedforward to work. Loading down the FET is the major no-no.

You can't keep a gain of 1.000 all the way to daylight, it's true.
However, the practical limit is set by the differentiated noise of the
first stage bootstrap device, in this case a single BF862. In my
world, once the SNR starts to tank, bandwidth isn't important any more.



I rely very heavily on capacitance multipliers in front ends. They're
squishy down at low frequency where nobody cares, but they're clean as a
whistle at high frequency where it matters. You can get something like
140-160 dB rejection at SMPS frequencies if you're willing to spend 1.5
volts or so, e.g. this one that I posted last year under the topic
"Improved capacitance multiplier". You really can't do high
performance front ends without them these days--quiet linear-regulated
supplies are nearly extinct.


Q1 Q2

IN DNLS160 DNLS160 3 OUT

0-----*--- --------- ------RR-*--------0
| \ / \ / |
| \ A \ A |
| --------- --------- |
390 R | | === 47uF alum
R | | | || 1uF X7R
| 390 | 330 330 | |
| | | GND
*--RR---*---RR---*---RR---*----+
| | | | |
| | | | R 91k
=== === === === R
1u | 1u | 1u | 1u | |
| | | | |
GND GND GND GND GND

Don't forget the reverse protection diode across the whole chebang and
reverse Vbe protection. There is always a chance that something else
shorts out the input rail and then the 47uF cap tries to feed things
upstream.

Oh, and the DNLS160 is kinda pricey

 

[Joerg]

Phil Hobbs wrote:

[...]

Don't forget the reverse protection diode across the whole chebang and
reverse Vbe protection. There is always a chance that something else
shorts out the input rail and then the 47uF cap tries to feed things
upstream.

Oh, and the DNLS160 is kinda pricey

Good points.

The one I posted above was for a 48V supply, which also gets a bit
unpleasant when you power it from a +48V lead acid battery--lots of
Joergish noises when you hit the switch. An inverted MOSFET in series,
with a bit of RC delay in the gate, fixes that pretty well. Shunting
the C-E of Q1 and Q2 with a couple of rectifiers in series each would
probably do just as well. In principle, you still need that series
diode to protect against input shorts

The one I'm using just now is for a +15V photodiode bias supply. The
higher current supplies are +-5, so although I would have liked to use a
cap multiplier, I didn't have the luxury. They had to be pi networks
with 180 uF aluminum polymer electros and 390 uH toroids (you want to
know about pricey...). Hopefully two poles is enough there.

If you don't need a lot of current, how about this series? Less than 20
cents:

http://www.tdk.co.jp/tefe02/e531_vlp8040.pdf

Bigger, about 40 cents:

http://www.bourns.com/data/global/pdfs/SRR1280.pdf

 

[Joerg]

<Big snip of a great thread!>

We could always start one about global warming

Reverse Vbe protection! Ouch!

I think that may be what 'took out' a transitor in a fault I could
never reproduce.
(I designed in a replacable transitor, which is a bit of a cludge.)
Some arc's 'n sparks tommorrow now that I have a 'theory' as to the
cause.

Thanks Joerg.

You are welcome. Those are the things that can sneak in via the back
door. Very nasty when defects only occur once in a blue moon and none of
the operators can remember what they've done differently than usual.

[...]

 

[Joerg]

On 09/19/2011 01:46 PM, Joerg wrote:
Phil Hobbs wrote:
[...]


Q1 Q2

IN DNLS160 DNLS160 3 OUT

0-----*--- --------- ------RR-*--------0
| \ / \ / |
| \ A \ A |
| --------- --------- |
390 R | | === 47uF alum
R | | | || 1uF X7R
| 390 | 330 330 | |
| | | GND
*--RR---*---RR---*---RR---*----+
| | | | |
| | | | R 91k
=== === === === R
1u | 1u | 1u | 1u | |
| | | | |
GND GND GND GND GND




Don't forget the reverse protection diode across the whole chebang and
reverse Vbe protection. There is always a chance that something else
shorts out the input rail and then the 47uF cap tries to feed things
upstream.

Oh, and the DNLS160 is kinda pricey


Good points.

The one I posted above was for a 48V supply, which also gets a bit
unpleasant when you power it from a +48V lead acid battery--lots of
Joergish noises when you hit the switch. An inverted MOSFET in series,
with a bit of RC delay in the gate, fixes that pretty well. Shunting
the C-E of Q1 and Q2 with a couple of rectifiers in series each would
probably do just as well. In principle, you still need that series
diode to protect against input shorts

The one I'm using just now is for a +15V photodiode bias supply. The
higher current supplies are +-5, so although I would have liked to use a
cap multiplier, I didn't have the luxury. They had to be pi networks
with 180 uF aluminum polymer electros and 390 uH toroids (you want to
know about pricey...). Hopefully two poles is enough there.

If you don't need a lot of current, how about this series? Less than 20
cents:

http://www.tdk.co.jp/tefe02/e531_vlp8040.pdf

Bigger, about 40 cents:

http://www.bourns.com/data/global/pdfs/SRR1280.pdf

It's going to hang off a large vacuum chamber with a 50 kW (average!)
pulsed CO2 laser and a bunch of motors and other stuff, and it's
providing the measurement data that everything else depends on, so I
really want toroids. Nobody will notice the extra $5, but they'd sure
notice if it were vulnerable to pickup. I'm really tempted to put in a
two-stage pi network, just to get another couple of poles.

If pickup is super-critical here is an old trick:

Assume the toroid core is ideal and keeps the field 100% in. Then you
can still have one effective turn in air (the overall "turn" that
follows the core) unless the winding trudges back to where it begun (but
that in turn can result in unwanted capacitive coupling). Sometimes
running the wire back as a loop can cancel most of that. Needs to be
"experimented out".

Sometimes winding your own is better than buying. Had to do that on an
RF pulser prototype late last week (after my right front paw had healed
from poison oak).

 

[Joerg]

George Herold wrote:

[...]

Hmm, It was myself and not an operator that blew it up. There are a
bunch of low noise power suppliles, with current limiting at the
input. But behind that, nothing but transistors. I 'tested' the
supplies by shorting 'em all, but there very well may have been a big
cap on one of the supplies when I blew it. I never tested for that!

3rd party power supplies? Yeah, I really hate that. They never state a
value for the output cap in their datasheets. That cap is usually behind
any and all current limiting so if a short happens its energy is going
to be dumped into your circuit at full gusto.

So when I get a new I often open it to see what and how much they've got
in there.

 

[Joerg]

On 09/19/2011 07:56 PM, Joerg wrote:
Phil Hobbs wrote:
On 09/19/2011 01:46 PM, Joerg wrote:
Phil Hobbs wrote:

[...]


Q1 Q2

IN DNLS160 DNLS160 3 OUT

0-----*--- --------- ------RR-*--------0
| \ / \ / |
| \ A \ A |
| --------- --------- |
390 R | | === 47uF alum
R | | | || 1uF X7R
| 390 | 330 330 | |
| | | GND
*--RR---*---RR---*---RR---*----+
| | | | |
| | | | R 91k
=== === === === R
1u | 1u | 1u | 1u | |
| | | | |
GND GND GND GND GND




Don't forget the reverse protection diode across the whole chebang
and
reverse Vbe protection. There is always a chance that something else
shorts out the input rail and then the 47uF cap tries to feed things
upstream.

Oh, and the DNLS160 is kinda pricey


Good points.

The one I posted above was for a 48V supply, which also gets a bit
unpleasant when you power it from a +48V lead acid battery--lots of
Joergish noises when you hit the switch. An inverted MOSFET in
series,
with a bit of RC delay in the gate, fixes that pretty well. Shunting
the C-E of Q1 and Q2 with a couple of rectifiers in series each would
probably do just as well. In principle, you still need that series
diode to protect against input shorts

The one I'm using just now is for a +15V photodiode bias supply. The
higher current supplies are +-5, so although I would have liked to
use a
cap multiplier, I didn't have the luxury. They had to be pi networks
with 180 uF aluminum polymer electros and 390 uH toroids (you want to
know about pricey...). Hopefully two poles is enough there.


If you don't need a lot of current, how about this series? Less than 20
cents:

http://www.tdk.co.jp/tefe02/e531_vlp8040.pdf

Bigger, about 40 cents:

http://www.bourns.com/data/global/pdfs/SRR1280.pdf


It's going to hang off a large vacuum chamber with a 50 kW (average!)
pulsed CO2 laser and a bunch of motors and other stuff, and it's
providing the measurement data that everything else depends on, so I
really want toroids. Nobody will notice the extra $5, but they'd sure
notice if it were vulnerable to pickup. I'm really tempted to put in a
two-stage pi network, just to get another couple of poles.

If pickup is super-critical here is an old trick:

Assume the toroid core is ideal and keeps the field 100% in. Then you
can still have one effective turn in air (the overall "turn" that
follows the core) unless the winding trudges back to where it begun (but
that in turn can result in unwanted capacitive coupling). Sometimes
running the wire back as a loop can cancel most of that. Needs to be
"experimented out".

Sometimes winding your own is better than buying. Had to do that on an
RF pulser prototype late last week (after my right front paw had healed
from poison oak).

Thanks. The ones I'm using have the self-reversed winding, where the
winding crosses over a core diameter and then keeps going, to get rid of
most of the solenoidal field.

Just crossing the core in a bee-line may not provide 100% compensation.
Nothing ever will but it can pay to bend and experiment. Sometimes you
cna get another 3dB for free.

 

[Joerg]

George Herold wrote:

[...]

Hmm, It was myself and not an operator that blew it up. There are a
bunch of low noise power suppliles, with current limiting at the
input. But behind that, nothing but transistors. I 'tested' the
supplies by shorting 'em all, but there very well may have been a big
cap on one of the supplies when I blew it. I never tested for that!

3rd party power supplies? Yeah, I really hate that. They never state a
value for the output cap in their datasheets. That cap is usually behind
any and all current limiting so if a short happens its energy is going
to be dumped into your circuit at full gusto.

So when I get a new I often open it to see what and how much they've got
in there.

No, the power supply is all mine. (Of course pieces copied from
everywhere.)
A cap multiplier as Phil posted, but only one transistor.

I may have more questions tomorrow, if I can get it to blow again.

Please don safety googles when trying that

Seriously, I've seen a close call once.