Preamp for Analog Discovery 2

[thanasisk]

I am investigating the use/design of a preamp for measuring low-level (<10mV) low-frequency (<10 MHz) signals with the Analog Discovery 2. The idea is to boost the measured signals above the noise floor of the unit.
The AD2 has an Input impedance of 1MOhm || 24pf but unfortunately the sensitivity is +-2.5V full-scale (500mV/div) in “high gain mode”.

Theoretically a 1000x(60dB) gain could yield 500uV/div.
Or a 250x gain for 2mV/div.

In general I found numerous preamp designs based on op-amps and jfets+op amps and paralleled op-amps and other mic preamp evaluation boards (from thatcorp). They offer excellent noise+THD performance BUT they all have been specifically designed for low source impedance (1KHz..10KHz).

In which case, if  a 10x probe is used at the input or if the test point is high impedance, then the accuracy of measurement due to loading and the noise & distortion of the preamps would severely suffer.

The Renesas ISL28634EV2Z also looks nice (differential-in differential-out) but the gain-bandwidth of the instrumentation amplifier used is not that great (1000x at up to 1 KHz approx.)

- Any thoughts/ideas/recommendations?
- If I have to design something from scratch, do you think a single low-noise op-amp stage (or instrumentation amp) with 60dB gain at >20KHz be feasible?
- Also am I right to look for op-amps which perform well with high source impedance of approx. 10MOhm to support a 10x probe and/or high impedances for probing with a 1x probe? See Fig. 2 of  http://www.analog.com/media/en/reference-design-documentation/design-notes/dn355f.pdf
- Or alternatively would you recommend to look at discrete jfet designs? Such as Fig. 4 from http://www.janascard.cz/PDF/Design%20of%20ultra%20low%20noise%20amplifiers.pdf

[rstofer]

Dave has a video on extending bandwidth by limiting the gain of each stage.  So, more than one stage...



Dave also has the uCurrent gadget.  You can search for it but here is the schematic

http://www.eevblog.com/files/uCurrentRev5schematic.pdf

[JS]

AD620 might be a good option, ~90k source 20dB seems to be the sweet spot for noise (not for offset) and then you can go for a few extra stages with opamps, let them clip, maybe with some diodes to just select between different stages for the sensitivity señection. If you use higher rails you could use the attenuator in the analog discovery to select between intermediate steps.

JS

 If using an instrumentation amp you could take the go to make it differential

[EmmanuelFaure]

Some ideas :
AD8428 : Bandwidth 3.5 MHz at fixed gain = 2000. But not adapted to high source impedance (Bipolar input with high bias current and current noise).
LT1102 : Bandwidth 3.5 MHz at fixed gain = 10. JFET input, so well adapted to high source impedance.

[cat87]

How about this? 

http://techlib.com/electronics/audioamps.html

Also,  if you want to go the in-amp route,  maybe the INA111 or INA112?

[DaJMasta]

If you find something you like that needs a low source impedance, you can always put a high impedance buffer as a frontend.  If your buffer stage is either unity gain or low gain, it doesn't need nearly the gain bandwidth product as your main amp, so it should be easier to find something that's low noise and has very little current requirements from your probe.  I definitely think cascading amps is going to be your best bet if you're needing several MHz of bandwidth and very low noise - most of the amps with several hundred MHz GBP required to get good gain and frequency response in one stage generally have higher noise figures near DC (audio amps are often spec'd at 1kHz or 10kHz whereas faster amps are spec'd at 1MHz and the noise figure generally goes up exponentially towards DC), so while they'd be fine as a high gain stage for a reasonably fast AC signal, you're going to lose the low level fidelity near DC.  Using amps designed more for low frequency use will give you much better noise performance (and when looking at very small signals, this will be important), but will generally have much lower gain bandwidth product, giving you much lower maximum gain for a given bandwidth requirement.


It's also probably worth including a low pass filter for your intended maximum bandwidth in your frontend - reducing the overall bandwidth of the system inherently reduces the noise floor, so limiting it to the frequency span you want to look at will improve low noise performance.  Looking at the very small signals also means you're much more susceptible to grounding issues, supply noise, and EMF.... so be careful the way you layout the design.  Local regulation, ample decoupling, an analog ground plane with a wide single path to ground, and maybe even shielding could be beneficial to the overall performance of the design.

[tggzzz]

Don't forget to calculate the unavoidable noise (shot, thermal) in your front-end.

[David Hess]

- Any thoughts/ideas/recommendations?

Check out the Tektronix AM502 which can provide a gain of 100,000 (100db) with a 1 MHz bandwidth.  It was made for exactly this type of application.

- If I have to design something from scratch, do you think a single low-noise op-amp stage (or instrumentation amp) with 60dB gain at >20KHz be feasible?

Not without a lot of care in design and construction.

- Also am I right to look for op-amps which perform well with high source impedance of approx. 10MOhm to support a 10x probe and/or high impedances for probing with a 1x probe?

There are some suitable JFET input operational amplifiers but watch out for variation of input bias current and input capacitance with signal level.  Low frequency noise will be a limitation and MOSFETs are worse in this regard.  High open loop bandwidth in the first stage to get high gain will be counterproductive because high bandwidth devices are also higher noise.

- Or alternatively would you recommend to look at discrete jfet designs?

Check out the Tektronix AM502 for an example of a discrete JFET design.  Notice that neutralization was used and gain of the JFET stage is relatively low.

A discrete JFET input design could have 1/4 the noise of an equivalent integrated JFET design but other considerations are more important.

Such as Fig. 4 from http://www.janascard.cz/PDF/Design%20of%20ultra%20low%20noise%20amplifiers.pdf

That is AC only.

I suspect you are going to want a differential input like the AM502 to avoid noise contributed by ground loops.

[JS]

...
The AD2 has an Input impedance of 1MOhm || 24pf but unfortunately the sensitivity is +-2.5V full-scale (500mV/div) in “high gain mode”.

Theoretically a 1000x(60dB) gain could yield 5uV/div.
Or a 250x gain for 2mV/div.
...

You forgot a few zeros somewhere, 500mV/1000 is 500µV, not 5µV. How low is the noise from the AD2? I guess you would want to match that noise or there about, there's no reason to go much under as you will get the same noise, and there's no reason to get much higher or you are amplifying more than it's useful.

Still, not very low noise. If you are ok with 1MHz bandwidth it's not that hard, many options for low noise, 10MHz opamps, using a gain up to 10x on each stage and you will be fine. Using the OP27 as first stage will give noise floor for 8 bit resolution at 60dB final gain @ 1MHz BW and be good down at DC. If you implement a selectable filter you can get even better. Going up to 10MHz is a bit harder as low noise opamps are harder to find and more expensive, and close to DC aren't as good anymore.

JS

[cat87]

Just throwing this in the pile, maybe it'll be useful to you at some point.

If you want to go all the way down to DC and up to say 10 or 20 Hz,  you can use a split path amp. You can have you ac-coupled amp,  say 10Hz to 1MHz  then,  also have something like a LTC1050  handling the DC part,  in parallel. Identical gain,  of course.

There's something  regarding this in App Note 106 fron  Linear.

[mtdoc]

Check out the Tektronix AM502 which can provide a gain of 100,000 (100db) with a 1 MHz bandwidth.  It was made for exactly this type of application.

+1 for the AM502. The selectable low and high frequency filters are a boon as well.

[thanasisk]

Thank you all for the excellent pointers!

Yes of course I meant 500uV/div 

The AM502 looks nice but needs a TM501 power module if I am not mistaken to go with it. I have seen listimgs on ebay but, because of the age of the devices, it is completely unclear whether any sold device performs according to spec.

I am not sure about the noise floor of the AD2 since I do not currently have the device and all measurements by other users online were not terminated properly. Of course this is essential info needed to set the design goals for this project. Any user of AD2 out there care to measure the AD2 noise floor?

 I realize that a nicely optimized two stage cascade would be the best option. A high impedance low noise  1x or 10x stage followed by a higher gain stage.

Since the inputs of the AD2 are differential, a cascade of diff in diff out instrumentation amps would be ideal. And a switchable bandpass or low pass filter would further limit the noise.

That is a lot of stuff to study and there are a lot of options to consider.. 

By the way, I also saw this evaluation board which offers 60db gain at sufficient BW

http://www.analog.com/en/design-center/evaluation-hardware-and-software/evaluation-boards-kits/eval-ad8253.html#eb-documentation

But I cannot find it in Europe..

[xani]

I am not sure about the noise floor of the AD2 since I do not currently have the device and all measurements by other users online were not terminated properly. Of course this is essential info needed to set the design goals for this project. Any user of AD2 out there care to measure the AD2 noise floor?

If you tell me what exactly do you need, I can do it, I'm not an expert on low noise measurement so I assume you need a bit more than just to short output and measure that ?

I was toying with similar idea as yours, I wanted to make a "front panel" with a bunch of switches (basically nicer version of AD2's BNC adapter) and possibly a front end for it.

[JS]

I don't kbow the optimal source impedance of the ad2, but having a noise measurement with shorted input and maybe 2 or 3 with other impedances would do. 1k, 10k, 1M would be my choices in the dark, from those thw optimal could probably be extracted. Just a resistor in the input, lower range, rms and peak to peak will give more than enough information. The knee for 1/f would be sweet but I don't think it's essential.

JS

[xani]

From schematics it seems to just have ~1M input with a resistor divider followed by analog switch and FET amp so it is likely that input impedance won't matter much.

I'll try to get it measured this evening, provided today's day at work wont have more suprises 

[David Hess]

There is no reason to match the impedance when a high input impedance buffer is used; by the time you have a low enough impedance to drive the input capacitance at the bandwidth you require, noise considerations are insignificant and the only thing which matters is the input voltage noise which can be calculated from the AD2 schematics.  Broadband noise starts at about 33nV/SqrtHz because of the minimum input divide by 5 stage.

[JS]

I was treating it as a black box, haven't look the schem...

33nV/sqrt(hz) then the AD620 has 13nV/sqrt(hz) so you only get a gain of 3 not good.

JS

JS

[thanasisk]

My recommendation for the noise measurement would be to use the BNC adapter board (and not the flywires) and to terminate the input with any BNC terminator you have available. Any terminator: short circuit, 50Ohm, 75Ohm or whatever you may have available will not affect the noise measurement at all because the equivalent source resistance for the input buffer will not be severely altered.

The AD8066 input buffer used has a corner frequency of around 2kHz (see fig. 21 in the datasheet). I believe this is the noisiest component in the chain.

[xani]

Okay, so termination are just 2 jumpers directly on AD2 socket. Last one is 1MHz sine fed from SDG1025 via BNC adapter

[JS]

Pictures in tapatalk sucks, really!
I'll need to come back in a few days on my pc, but I will put some effort into this as I do want to build a preamp of similar specs for my setup...

JS

[David Hess]

The AD8066 input buffer used has a corner frequency of around 2kHz (see fig. 21 in the datasheet). I believe this is the noisiest component in the chain.

The breakpoint for thermal noise in the input divider and series protection is higher than the 1/f corner frequency of the operational amplifier.  Assuming I did it right, below 320kHz noise rises to 12nV/SqrtHz and below 7.3kHz, noise rises to 55nV/SqrtHz so the 1/f corner frequency is actually about 30Hz.  Most oscilloscope inputs avoid this level of noise by not using an input divider on their most sensitive ranges and including a lot more bypassing on the series protection resistor although this makes input protection more difficult.

As a practical matter, this means excluding quantization noise, the external preamplifier should have enough gain to overwhelm 55nV/SqrtHz if performance below about 7.3kHz matters and that is the design constraint I would start with.  For an integrated JFET amplifier, that means at least a gain of about 10 minimum.  For a lower noise discrete JFET, that means a gain of about 50 minimum.

Update: Those were the noise values at the input to the operational amplifier.  The input referred noise for the AD2 is 5 times higher because of the attenuator so 60nV/SqrtHz below 320kHz and 275nV/SqrtHz below 7.3kHz.  So for an integrated JFET amplifier, that means at least a gain of about 50 minimum and for a lower noise discrete JFET, that means a gain of about 250 minimum.

[Krytron]

Look at the JFET designs:  internet search "Design of Ultra Low Noise Amplifier, Vojtech Janasek, www.janascard.cz" under downloads (on the main page to the right is button for english).

[thanasisk]

Thank you all!
I ll try to come up over the weekend with a rough first design with differential inputs and outputs using instrumentation amps using all this info here..
Later i can look at more elaborate discrete designs..

I am wondering whether the IN+, IN- and GND inputs should be BNC connectors for connecting typical oscilloscope probes..  Does it make sense? I saw in a design by Walt Jung (In regulators for high-performance audio, 1995) that he used shielded twisted pair for direct probing.

[xani]

Well if you wanna go all in you could possibly make an active probe-type of solution with head that has say x100 amplifier sending that to the rest via twisted pair or BNC.

But I'd say test with just a normal BNC + scope probes and only bother upgrading from that if that is not enough. Also there is always option of just soldering BNC cable directly to the target

[thanasisk]

Ok so Xani I should be looking at the figure showing approx. 900uVrms (5.8mVpp) noise, I suppose this corresponds to the "high gain mode" of the AD2. This noise results in SNRFS (5Vpp full-scale sin signal) of approx. 65.86dB -> ENOB of 10.6 bits. (By the way your results are similar to these: https://forum.digilentinc.com/topic/4306-analog-discovery-2-wavegen-noise-below-50mv/)


Thanks David for the very detailed pointers. They were very useful to understand what is going on with the noise calculations.


Now lets see how a single stage differential-single ended preamp based on the LT1102 (input referred noise voltage of 20nV/sqrt(Hz)) could perform:

- One LT1102 configured as G=10 (3.5MHz BW) would offer a 50mV/div sensitivity to the AD2. Output referred voltage noise of 200nV/sqrt(Hz) would mean approx 374uVrms@3.5MHz BW. However, performance below 7.3kHz would suffer because the gain is not enough to push over the AD2 noise floor in that region.

- One LT1102 configured as G=100 (220kHz BW) would offer a 5mV/div sensitivity to the AD2. Output referred voltage noise of 2000nV/sqrt(Hz) would mean approx 938uVrms@220kHz, offering 10.6 bits ENOB.


A differential out configuration similar to p. 9 of the datasheet would double the gains to 20/200 but suffer from increased voltage noise.

- Two LT1102 configured as G=20 (3.5MHz BW) would offer a 25mV/div sensitivity to the AD2. Output referred voltage noise is 566 nV/sqrt(Hz).

- Two LT1102 configured as G=200 (220kHz BW) would offer a 2.5mV/div sensitivity to the AD2. Output referred voltage noise is 5657nV/sqrt(Hz).

Edit: updated the noise calculations for the differential output case

[David Hess]

The differential configuration assuming you drive two channels of the AD2 raises the AD2's input noise by Sqrt(2).  Differential amplifiers like the 7A13 and 7A22 have more noise even when one input is grounded compared to single ended amplifiers.

Of course the same thing applies when two single ended channels are used in add and invert mode except if this is done digitally, then the quantization noise of the separate channels *also* adds together.

[xani]

Ok so Xani I should be looking at the figure showing approx. 900uVrms (5.8mVpp) noise, I suppose this corresponds to the "high gain mode" of the AD2. This noise results in SNRFS (5Vpp full-scale sin signal) of approx. 65.86dB -> ENOB of 10.6 bits. (By the way your results are similar to these: https://forum.digilentinc.com/topic/4306-analog-discovery-2-wavegen-noise-below-50mv/)

Also note that both channels are shifted slightly from zero, and by different amounts.

As for design, you can do something clever like auto-switching gain of input amplifier via software using digital IO pins. I've seen some clever scripts written for it, like THD meter that just scripted built-in instruments to graph THD/frequency

[thanasisk]

The differential configuration assuming you drive two channels of the AD2 raises the AD2's input noise by Sqrt(2).  Differential amplifiers like the 7A13 and 7A22 have more noise even when one input is grounded compared to single ended amplifiers.

Of course the same thing applies when two single ended channels are used in add and invert mode except if this is done digitally, then the quantization noise of the separate channels *also* adds together.

The differential output is used to drive a single differential channel of the AD2.

I am trying to understand the operation of the circuit of p. 9 (fig. on the top left) of the datasheet. There are two diff-in single-ended-out LT1102 amplifiers connected in (inverted) parallel mode. The positive output is the normal signal amplified by the top amplifier (with G=10 or 100) plus the output referred voltage noise of the top amplifier. The negative output is the inverted signal amplified by the bottom amplifier (with G=10 or 100) plus the inverted output referred voltage noise of the bottom amplifier. So the diff output is G*Vi+G*N1-G*(-Vi)-G*N2 = 2*G*Vi + G*(N1+N2). The output referred voltage noise is then INCREASED by sqrt(2) and the voltage gain is G=20..200 depending on the config of the top&bottom LT1102s.

By the way, the diff output configuration in the LT1102 data sheet p.9 is very similar to p. 35 of http://www.thatcorp.com/datashts/More_Analog_Secrets.pdf

There it also says that a "low-Z attenuator" between the (differential) output and the ADC input is required: "optimizes noise & headroom". Is  this statement importance for our case (whereas the AD2 input attenuator is high-Z)?

Edit: corrected the output noise conclusion for the parallel inverted case..

[thanasisk]

Ok so Xani I should be looking at the figure showing approx. 900uVrms (5.8mVpp) noise, I suppose this corresponds to the "high gain mode" of the AD2. This noise results in SNRFS (5Vpp full-scale sin signal) of approx. 65.86dB -> ENOB of 10.6 bits. (By the way your results are similar to these: https://forum.digilentinc.com/topic/4306-analog-discovery-2-wavegen-noise-below-50mv/)

Also note that both channels are shifted slightly from zero, and by different amounts.

As for design, you can do something clever like auto-switching gain of input amplifier via software using digital IO pins. I've seen some clever scripts written for it, like THD meter that just scripted built-in instruments to graph THD/frequency

This "DC-RMS" measurement, this is the DC offset of the channels right? I suppose this reduces the dynamic range so in my calculations for the SNR in the 500mV/div "high-gain" range of the AD2, I should take the maximum input signal as 5Vpp*0.3535 (=Vrms)  minus this DC-RMS offset?

Yes indeed a switching system with e.g. relays using the digital IO pins is a cool idea! I believe it is used with the impedance analyzer board.

[David Hess]

The differential configuration assuming you drive two channels of the AD2 raises the AD2's input noise by Sqrt(2).  Differential amplifiers like the 7A13 and 7A22 have more noise even when one input is grounded compared to single ended amplifiers.

Of course the same thing applies when two single ended channels are used in add and invert mode except if this is done digitally, then the quantization noise of the separate channels *also* adds together.

The differential output is used to drive a single differential channel of the AD2.

I am trying to understand the operation of the circuit of p. 9 (fig. on the top left) of the datasheet. There are two diff-in single-ended-out LT1102 amplifiers connected in (inverted) parallel mode. The positive output is the normal signal amplified by the top amplifier (with G=10 or 100) plus the output referred voltage noise of the top amplifier. The negative output is the inverted signal amplified by the bottom amplifier (with G=10 or 100) plus the inverted output referred voltage noise of the bottom amplifier. So the diff output is G*Vi+G*N1-G*(-Vi)-G*N2 = 2*G*Vi + G*(N1+N2). The output referred voltage noise is then reduced by sqrt(2) and the voltage gain is G=20..200 depending on the config of the top&bottom LT1102s.

Isn't my analysis correct?

I do not think that is strictly correct.

The output voltage is doubled but the noise in each amplifier is uncorrelated with the other so it adds as well when seen at the output; it is like having the two noise sources in series resulting in Sqrt(2) higher noise at the output.  This is still an improvement in signal to noise ratio since the signal gain is even higher.

But compared to a single ended amplifier, the differential configuration has more noise and a lower signal to noise ratio.

By the way, the diff output configuration in the LT1102 data sheet p.9 is very similar to p. 35 of http://www.thatcorp.com/datashts/More_Analog_Secrets.pdf

That is a nice reference from THAT Corp.  It is the same circuit.  Using a pair of instrumentation amplifiers like that has the advantage of providing high impedance inputs.  "Op Amp Applications Handbook" by Walter Jung from Analog devices gives some examples of similar circuits starting on page 480 and 520.

There it also says that a "low-Z attenuator" between the (differential) output and the ADC input is required: "optimizes noise & headroom". Is  this statement importance for our case (whereas the AD2 input attenuator is high-Z)?

It has nothing to do with the input attenuation.  The point they are making is that if the amplifier's output noise is significantly higher than the input noise of the next stage (ADC), then the excess gain has little benefit and only serves to lower the dynamic range.  Radio receivers face this problem all the time.

[blackdog]

Hi,

There are many ways to build a preamp and i want to show one of my design.
It is doing 60dB, has reasonable low noise and a good puls response.
More than 6MHz -3dB point, about 4MHz -1dB and 900KHz -0.1dB.
Square wave has no overshoot :-)   (If you build the circuit correctly)




The amplifier has input and output protection and can work on 2x9V battery or if you need more output on a mains power supply, but watch for the power dissipation of the THS3061!
I hackt the THS3061 with a heathsink and some thermic glue to get a higer output when working on a 2x12V power suppply.
The square-wave response is without abberations is only possible if the connections are kept short!
And maybe adjust the capacitor C25 a little.

Build to work directly connected to a D.U.T or via a scope probe.
The output is designed to drive a NOT terminated schort peace of coax cable (Like a RG58 BNC patch kabel not more than 3 feet)

Dutch website with explaning and picture, Google translate is your friend.
https://www.circuitsonline.net/forum/view/135023#highlight=meetversterker


PS,
No, you can't replace the opamps and then expect that the circuit will give the same performance., changing something and you are on your own :-)

Kind regarts,
Bram

[thanasisk]

Oh yes, of course the uncorrelated noise adds up! Thank you for the acute observation and also the very interesting reference David. I have updated my nose calculations in my posts above. The output noise for a differential output with two LT1102s in parallel inverted mode is indeed quite higher while I suppose the performance gains achieved (CMR rejection,..) compared to a single ended solution would not pay off (we will not have to drive any significant cable/pcb trace length at the output anyway.. ).  Also given the price of one LT1102, makes me wonder whether it is better to drop the differentential output requirement altogether or just use a secondary cascaded stage with e.g.THAT1606 or similar (have to
do the noise calculations for a cascade as well..)

Bram thanx for the interesting design, I found it before when I was researching.  My requirements for this design is high input impedance  with differential inputs (at least at this stage)..

[thanasisk]

As a side note, I was amazed at the scarcity of high input impedance amps offered out there.. Makes design life easier i guess by not having ample choices 

[blackdog]

Hi thanasisk, :-)

Misschien zoek je naar de verkeerde de opamp?    (Maybe you are looking for the wrong the opamp?)

The problem is that the fast opamps almost all have a low input impedance, but you were already familiar with that.
There is a beautiful TI IC that has a Fet input and that might have a low enough noise number for you.
The gain is 2x of this IC and you can then use fast BJT opamps behind for your gain of 40 or 60dB.
You will have to calculate if the 1/F noise is low enough for you.
Maybe more than one parallel if you just don't get your noise figure...
http://www.ti.com/product/opa653/description?keyMatch=opa653&tisearch=Search-EN-Everything

Kind regards,
Bram

[xani]

Also given the price of one LT1102, makes me wonder whether it is better to drop the differentential output requirement altogether or just use a secondary cascaded stage with e.g.THAT1606 or similar (have to
do the noise calculations for a cascade as well..)

AD6066, which is used in AD2, looks pretty interesting (FET, low noise), but probably bit too slow to get to the full AD2 speed with any significant gain. But it is a bit cheaper than LT1102.

[thanasisk]

Both AD8066 and OPA653 are interesting parts but my understanding is that they were specced for high speeds&bandwidth at the expense of higher noise in low frequencies (and I do want decent  DC-22KHz performance..) .  Also I would be reluctant to consider them as they do not seem to be very forgiving to a diyer in terms of pcb design and parasitics.. 

[blackdog]

Hi thanasisk,

Really very low noise such as for a good dynamic microphone amplifier or a band microphone is something quite different from a low noise amplifier with an input impedance of 1Meg.
You have to decide what you want, whether you want it or not. 

You can build an amplifier with a Fet input like with the 2SK170 and its family,
but it will always have a large input capacity and the bandwidth of that amplifier will depend on your steering-impedance.
The latest version of The Art Of Electronics contains a few amplifiers with graphs of the bandwidth to be achieved with respect to the input impedance.

To measure low noise of power supplies and voltage references, you do not need a 1Meg input impedance, 10 to 20K is usually sufficient.
It usually comes down to the fact that you will need more than one type of preamplifier for your LAB.

Then again, don't just think of amplifiers, but also of filters!
Often it is useful to have some filters available after the first amplifier stage so that e.g. hum can be removed via a notch filter without overloading your measuring device.
A second handy filter is 400Hz High Pass to filter everything below that frequency.
Think of 12 to 18dB per octave for the filters and 50dB for the notch.

Kind regards,

Bram

[thanasisk]

The main idea behind this design was to extend the capabilities (sensitivity) of the AD2 without altering its "character" (1M input impedance and differential inputs).

Very low noise low input impedance preamps are very tempting  but a high input impedance is really what I need for also being able to support a 10x passive probe at the input.

Yes indeed adding filters are important (notch,HPF but LPF as well) and I will do consider adding them later in the design process..

[DaJMasta]

I still think high input impedance is a near trivial addition if you just have a low gain buffering stage.  Say you want 10MHz of bandwidth, so you find a high input impedance amp (maybe start looking at < 10nA input bias current) and a low noise figure.... but instead of looking the perfect frontend amp with 500MHz+ GBP (and a sky high price) to do all your gain in that stage... you look for 50MHz GBP or less.  Then your input amp is very high impedance, low noise, and offers some gain, maybe you find one with 30MHz GBP so you use a fixed gain of 3, which costs a tenth or less of the fancy high performance amp and offers lower noise.  The output of that amp then drives enough current to be able to power almost any amp for your primary gain stage, and your secondary amp can be higher noise as well - the noise floor will be at least the input amp's noise times its gain, so in this example, your second stage amp for high gain can be up to 3x the noise figure before it starts dominating the noise of the pair.

I don't know if you care about selective filtering, but it is a good idea to put in basic low pass filters into your amp stages, because then any additional GBP over what is needed (in the spec or beyond) doesn't amplify signals that are outside of your specified bandwidth and contribute to your overall noise.  This can be as simple as an LC filter in the path between the two amps (and maybe another after the second stage) or a good dielectric cap in parallel with the feedback resistor if you're using an inverting amplifier configuration - minimal parts addition and design concern, but lower noise.

[thanasisk]

I still think high input impedance is a near trivial addition if you just have a low gain buffering stage.  Say you want 10MHz of bandwidth, so you find a high input impedance amp (maybe start looking at < 10nA input bias current) and a low noise figure.... but instead of looking the perfect frontend amp with 500MHz+ GBP (and a sky high price) to do all your gain in that stage... you look for 50MHz GBP or less.  Then your input amp is very high impedance, low noise, and offers some gain, maybe you find one with 30MHz GBP so you use a fixed gain of 3, which costs a tenth or less of the fancy high performance amp and offers lower noise.  The output of that amp then drives enough current to be able to power almost any amp for your primary gain stage, and your secondary amp can be higher noise as well - the noise floor will be at least the input amp's noise times its gain, so in this example, your second stage amp for high gain can be up to 3x the noise figure before it starts dominating the noise of the pair.

I don't know if you care about selective filtering, but it is a good idea to put in basic low pass filters into your amp stages, because then any additional GBP over what is needed (in the spec or beyond) doesn't amplify signals that are outside of your specified bandwidth and contribute to your overall noise.  This can be as simple as an LC filter in the path between the two amps (and maybe another after the second stage) or a good dielectric cap in parallel with the feedback resistor if you're using an inverting amplifier configuration - minimal parts addition and design concern, but lower noise.

Would you thus propose a differential op amp instead of an instrumentation amp? (see my requirements above).

An LPF I think will be I included and as you say can be configured with minimal part additions.

[David Hess]

Is there a specific reason you want to use two channels of the AD2 in a differential configuration?  Why not simplify things and convert the differential input to a single ended output?

Is fast overload recovery a requirement?  Shunt feedback amplifiers have problems with this.

It is not explicitly listed above but is common mode suppression a requirement?  This only applies if you want to use two channels of the AD2 as a differential input.

The dual LT1102 achieves this as shown in the datasheet and THAT's application note but it is not particularly low noise and the bandwidth at a gain of 10 (20 when using 2 as shown in the figure) is barely above 1 MHz; pay attention to the undistorted output versus frequency graph which shows the full power bandwidth at different output levels.  The alternative instrumentation amplifiers from AD, the AD8429 and AD8421 are not any better in regard to full power bandwidth with about the same slew rates.

Now the full power bandwidth can be increased by adding gain *after* the instrumentation amplifiers which is easy to do.  Doing so removes the limit of full power bandwidth discussed above from the instrumentation amplifiers so small signal bandwidth is again what matters and there the AD8429 and AD8421 are much better than the LT1102 which is stuck at a minimum gain of 10.

But if you use the AD8429 and AD8421, then a high input impedance buffer or preamplifier is required and here is where a pair of fast low noise JFET input operational amplifiers can serve.

[DaJMasta]

If you need the channel to be fully differential, you can always just duplicate the circuit - you add some variability to the system, but given the bandwidth you can probably just make sure your gain resistors are tight tolerance and your opamps have low input offset voltage and you'll probably be fine so long as the two are colocated on a board or in the same enclosure.  If you need a differential input, you can duplicate the first low gain stage and then send each into a secondary amp to move to single ended and do your final gain stage with that.  If you want differential input into the analog discovery for whatever reason but are fine with single ended inputs, you can just have an inverting follower amp after the output stage to get a mirrored signal (though I don't think it's going to promote signal integrity unless you need big time EMF rejection between the output of the preamp and the input of the analog discovery).

Personally, I don't know why you'd need a differential signal for this application, but perhaps your situation can demand it.  I would certainly prefer two distinct channels to play with and just go into a differential mode in software, and at these frequencies and signal levels, unless your area is very noisy in terms of EMF (and if it is, you probably need to shield the analog discovery too), single ended should be able to give you plenty of signal integrity and responsiveness within the bandwidth you're looking at and will be cheaper to implement (more choices of parts, cheaper options, fewer parts).

[xani]

The main idea behind this design was to extend the capabilities (sensitivity) of the AD2 without altering its "character" (1M input impedance and differential inputs).

Very low noise low input impedance preamps are very tempting  but a high input impedance is really what I need for also being able to support a 10x passive probe at the input.

Why would you want to measure very low signals using 10:1 probe ?

There is always an option of having a bunch of switches/relays and just having 2 switchable input amplifiers with different characteristics but that would complicate it a bit.

[thanasisk]

Been busy with simulating the whole frontend up to the ADC driver (used TINA-TI with AD8065 pspice model from AD). I just wanted to have a more precise understanding of the noise and frequency performance of the AD2 before discussing more about the requirements; I will thus get to your latest questions in a bit.

BW is 31.5 MHz (so this is what the AD2 website means by "30 MHz+")  BUT  what they fail to disclose is that the  low frequency performance does suffer! See attached Bode diagram.

I will therefore need a slight DC-100KHz boost in the preamp to equalize that 1-1.5dB non-flat frequency response.

The input referred noise is higher than what you had calculated David.

[David Hess]

BW is 31.5 MHz (so this is what the AD2 website means by "30 MHz+")  BUT  what they fail to disclose is that the  low frequency performance does suffer! See attached Bode diagram.

I will therefore need a slight DC-100KHz boost in the preamp to equalize that 1-1.5dB non-flat frequency response.

To me that looks like it is caused by improper compensation of the input divider.  The corner frequency is about right and the AD2 only includes one compensation adjustment.  High impedance attenuators are greatly affected by the printed circuit board material which is not accounted for in a simulation.  Look up oscilloscope "hook".

The input referred noise is higher than what you had calculated David.

It was only a "good enough" estimate and I did not bother including 1/f noise.  I wanted to find the reasonably minimum gain to design for.  I am pleased it is as close as it was.

1/f noise is a big deal in DC precision applications and where most integrated oscilloscope front ends are terrible; it is difficult to control in a wide bandwidth input stage.  Dual path input amplifiers which use a divider before the DC amplifier are even worse (as the AD2 design shows) and make those old dual JFET in a totem-pole configuration used for input buffers look really good.  They even make the Tektronix 7A13 look good.

The Tektronix 7A13 is my reference for oscilloscope vertical amplifiers because it is so noisy.  But it isn't a noisy as a modern oscilloscope!  The 7A13 comes out as better than 20nV/SqrtHz.  Single ended JFET input amplifiers of that age are about 3nV/SqrtHz.  The 7A13 is much worse because it has differential inputs and all of the extra circuitry to support a +/-10 volt common mode input range without input dividers.

What you seem to want is a differential probe which supports using x10 passive probes.  I might have suggestions for an alternative design but it depends on one of the questions I asked:

Is there a specific reason you want to use two channels of the AD2 in a differential configuration?  Why not simplify things and convert the differential input to a single ended output?

[thanasisk]

BW is 31.5 MHz (so this is what the AD2 website means by "30 MHz+")  BUT  what they fail to disclose is that the  low frequency performance does suffer! See attached Bode diagram.

I will therefore need a slight DC-100KHz boost in the preamp to equalize that 1-1.5dB non-flat frequency response.

To me that looks like it is caused by improper compensation of the input divider.  The corner frequency is about right and the AD2 only includes one compensation adjustment.  High impedance attenuators are greatly affected by the printed circuit board material which is not accounted for in a simulation.  Look up oscilloscope "hook".

The input referred noise is higher than what you had calculated David.

It was only a "good enough" estimate and I did not bother including 1/f noise.  I wanted to find the reasonably minimum gain to design for.  I am pleased it is as close as it was.

1/f noise is a big deal in DC precision applications and where most integrated oscilloscope front ends are terrible; it is difficult to control in a wide bandwidth input stage.  Dual path input amplifiers which use a divider before the DC amplifier are even worse (as the AD2 design shows) and make those old dual JFET in a totem-pole configuration used for input buffers look really good.  They even make the Tektronix 7A13 look good.

The Tektronix 7A13 is my reference for oscilloscope vertical amplifiers because it is so noisy.  But it isn't a noisy as a modern oscilloscope!  The 7A13 comes out as better than 20nV/SqrtHz.  Single ended JFET input amplifiers of that age are about 3nV/SqrtHz.  The 7A13 is much worse because it has differential inputs and all of the extra circuitry to support a +/-10 volt common mode input range without input dividers.

What you seem to want is a differential probe which supports using x10 passive probes.  I might have suggestions for an alternative design but it depends on one of the questions I asked:

Is there a specific reason you want to use two channels of the AD2 in a differential configuration?  Why not simplify things and convert the differential input to a single ended output?

This is all very interesting info and it uncovered a whole new world for me as I am not very experienced with instrumentation design. I will try to include a source impedance in the simulator and also to play with the variable capacitor value at the input attenuator of yhe AD2 to see what i can get..

I will agree with you that a single ended output is the way forward, so i look forward to your alternative recommendation

[thanasisk]

Bode response should be flat according to

https://reference.digilentinc.com/reference/instrumentation/analog-discovery-2/reference-manual?redirect=1#figure_14

I will rerun the sim to validate this

[David Hess]

There was another differential probe discussion a few week ago which covers my suggestion below.

My previous experiences using instrumentation amplifiers for differential probing at high bandwidth has never been satisfactory because of common mode rejection and other issues but there is an alternative now other than a discrete design.  Linear Technology and Analog devices make some current feedback based "difference amplifiers" which do exactly what you want with good performance and they are fast but lack high impedance inputs.  Download the datasheets to see how they are used.  I have been considering them so included the Excel file with my notes below.

+/-5 Volt Common Mode Voltage Range:

LT1187   25MHz   Minimum Gain of 2   65nV/SqrtHz
LT1189   18MHz   Minimum Gain of 10   30nV/SqrtHz
LT1193   40MHz   Minimum Gain of 2   50nV/SqrtHz
LT1194   35MHz   Minimum Gain of 10   15nV/SqrtHz
LT6552   37.5MHz   Minimum Gain of 2   55nV/SqrtHz

+/-10 Volt Common Mode Input Voltage Range:

AD830   85MHz   Minimum Gain of 1   27nV/SqrtHz   2V Differential Input
AD8129   185MHz   Minimum Gain of 10   4.5nV/SqrtHz   0.5V Differential Input
AD8130   250MHz   Minimum Gain of 1   12.3nV/SqrtHz   2.5V Differential Input

If you want a differential output with common mode suppression, then I guess you could double up and reverse the inputs just like you did with the LT1102 and this doubles the gain or bandwidth.  I do not see any advantage to driving the AD2 differentially unless you absolutely must have the highest possible dynamic range and even then the improvement by doing so is small.

The AD8129 with a gain of 10 has a bandwidth of almost 200MHz which is not needed but that means it can be configured for higher gains like 50MHz at a gain of 20 and almost 20MHz at a gain of 50 and this has no effect on the common mode input range.  The gain is also easily adjustable with a single ended circuit so the potentiometer trimmer can be grounded on one side.

For high impedance inputs, a preamplifier or buffer will need to be added.  This could be a pair of JFET operational amplifiers configured as buffers or more likely a two operational amplifier JFET differential amplifier with enough gain for a minimum input noise.  It just has to overcome the noise of the difference amplifier.

That leaves one problem which has not been discussed yet when using x10 probes; they don't match producing poor common mode rejection.  Tektronix used to make special probe sets to solve this but a better way in this case would be to trim the 1 megohm input shunts and compensation networks for a dedicated set of probes.  Or you could try selecting a set of probes which match.  X1 probes do not have this issue.

Note that x10 probes effectively raise the input noise of the amplifier by the same amount sort of defeating the purpose if you want low noise at any cost.  The same thing happens with that fixed attenuator on the AD2 inputs.

X10 probes could push the common mode input voltage range out to +/-100 volts.  That is pretty impressive with a potential broadband input noise of 60nV/SqrtHz which is 30 times better than a Tektronix 7A13 under the same conditions although the 7A13 is higher bandwidth.  Note that noise is what limits sensitivity unless DC drift is high.  At low frequencies the 9M resistor in the probe increases noise and there is nothing to be done about that other than not use an attenuating probe.

I could not find any instructions for using the forum "table" feature.

[thanasisk]

To me that looks like it is caused by improper compensation of the input divider.  The corner frequency is about right and the AD2 only includes one compensation adjustment.  High impedance attenuators are greatly affected by the printed circuit board material which is not accounted for in a simulation.  Look up oscilloscope "hook".

Exactly! What I forgot to simulate was the drain and source on/off capacitance of each one of the connections of the ADG612 which add a total of 15pF (typical) of capacitance in parallel with the 5-20pF trimmer. Then I am just 3pF away from a flat bandwidth response and a properly compensated square wave and without attempting to model any parasitics of the pcb. I can confirm that, also in my simulation, even a 1pF change to the compensation cap in the high impedance attenuator is observable both on the frequency response and on the square wave oscilloscope (virtual) instrument.

What do you mean by oscilloscope "hook"? (obviously it is not the hook connectors for the probes..)

[David Hess]

What do you mean by oscilloscope "hook"? (obviously it is not the hook connectors for the probes..)

Getting Rid of Hook: The Hidden PC-Board Capacitance

You probably will not have a problem but it is important where high impedance dividers are used like on the AD2.

[thanasisk]

If you need the channel to be fully differential, you can always just duplicate the circuit - you add some variability to the system, but given the bandwidth you can probably just make sure your gain resistors are tight tolerance and your opamps have low input offset voltage and you'll probably be fine so long as the two are colocated on a board or in the same enclosure.  If you need a differential input, you can duplicate the first low gain stage and then send each into a secondary amp to move to single ended and do your final gain stage with that.  If you want differential input into the analog discovery for whatever reason but are fine with single ended inputs, you can just have an inverting follower amp after the output stage to get a mirrored signal (though I don't think it's going to promote signal integrity unless you need big time EMF rejection between the output of the preamp and the input of the analog discovery).

Personally, I don't know why you'd need a differential signal for this application, but perhaps your situation can demand it.  I would certainly prefer two distinct channels to play with and just go into a differential mode in software, and at these frequencies and signal levels, unless your area is very noisy in terms of EMF (and if it is, you probably need to shield the analog discovery too), single ended should be able to give you plenty of signal integrity and responsiveness within the bandwidth you're looking at and will be cheaper to implement (more choices of parts, cheaper options, fewer parts).

A roll-my-own instrumentation amp with single ended ouput would make much sense. I will definitely try this,also with david's pointers.  I read somewhere (cant remember exactly where) that common mode rejection performance is bounded by the tolerance of the matched resistors and even selecting 0.1% resistors does not work wonders in that aspect. But other design aspects would be more important anyway.

[thanasisk]

What do you mean by oscilloscope "hook"? (obviously it is not the hook connectors for the probes..)

Getting Rid of Hook: The Hidden PC-Board Capacitance

You probably will not have a problem but it is important where high impedance dividers are used like on the AD2.

Looks like a complex problem to address. I trust that they dimensioned  the variable capacitor sufficiently in the AD2 (5-20pf) to roughly compensate for all these effects. As a side note, AD2s come precalibrated from the factory but it could be that they would also benefit from recalibration every now and then, especially after severe environmental  changes or aging.

[thanasisk]

8

Is fast overload recovery a requirement?  Shunt feedback amplifiers have problems with this.

Fast overload recovery is not a requirement (how fast are we talking here?) But I have always been wondering whether there exists is a distortion free, noise free and "cheap" way to indicate overload  with e.g. an led or digital output..

[David Hess]

I read somewhere (cant remember exactly where) that common mode rejection performance is bounded by the tolerance of the matched resistors and even selecting 0.1% resistors does not work wonders in that aspect. But other design aspects would be more important anyway.

DC and low frequency AC common mode rejection can be handled but high frequency AC common mode rejection is a major problem which involves both amplitude and phase.  Those parts I listed do not rely on matched impedances in a feedback network so have much better high frequency AC common mode rejection than would normally be the case.  High performance differential probes use a similar design.

Getting Rid of Hook: The Hidden PC-Board Capacitance

You probably will not have a problem but it is important where high impedance dividers are used like on the AD2.

Looks like a complex problem to address. I trust that they dimensioned  the variable capacitor sufficiently in the AD2 (5-20pf) to roughly compensate for all these effects. As a side note, AD2s come precalibrated from the factory but it could be that they would also benefit from recalibration every now and then, especially after severe environmental  changes or aging.

Compensating for hook in the circuit is possible but complex and hook often changes with humidity so this is not always feasible; the better option is to use a printed circuit board substrate which does not suffer from hook.  Before good FR4 substrates became available, Tektronix used special materials like polysulfone and an unknown white plastic substrate for their high impedance attenuators.

Fast overload recovery is not a requirement (how fast are we talking here?) But I have always been wondering whether there exists is a distortion free, noise free and "cheap" way to indicate overload  with e.g. an led or digital output..

The 7A13 uses feedback from the output to the input to clamp overload improving recovery time which could certainly be detected; the feedback only occurs when overload is present.  Monitoring the common mode and differential levels at the right points is enough to detect overload.

Overload is sneaky because not only does it "wind up" the integration stage used in feedback amplifiers, it also drives transistors into saturation or cutoff resulting in temperature changes which takes microseconds to milliseconds to resolve.

[PartialDischarge]

I wonder to what extent hook is still a problem with todays pcb manufacturing methods. Maybe nothing has changed from the 70s but modern literature on this topic seems inexistent.
Also the article mentions teflon standoffs as a way to support critical components whereas they don't mention slot cuts, todays CNCs make them in any shape and width, that is my preferred method of 'isolating' high impedance points.
The article also doesn't mention how the orientation of components in the board my affect performance in some cases, specially in HV circuits, this has to do with the orientation of fibers in the PCB

[David Hess]

I wonder to what extent hook is still a problem with todays pcb manufacturing methods. Maybe nothing has changed from the 70s but modern literature on this topic seems inexistent.

Nothing has changed for FR4 substrates and hook is still a problem; it is just that most circuits are not susceptible so nobody notices.  There have been reports of high voltage differential probes drifting out of their specifications which seem suspiciously like a problem with hook.  I have noticed it in high impedance circuits which do not settle as quickly as they should or mysterious non-linearity in high resolution analog to digital converters.  I suspect the designers of multimeters run across it in their high impedance dividers leading to errors in AC measurements at different frequencies.

Where the lack of hook matters, it is necessary to either qualify board manufacturers or use substrates which are guaranteed to have a low level of hook.

Also the article mentions teflon standoffs as a way to support critical components whereas they don't mention slot cuts, todays CNCs make them in any shape and width, that is my preferred method of 'isolating' high impedance points.

Usually the concern is controlling leakage.

The article also doesn't mention how the orientation of components in the board my affect performance in some cases, specially in HV circuits, this has to do with the orientation of fibers in the PCB

I know this is important for controlling the impedance of transmission lines at high frequencies on woven substrates but why would it matter for high voltage circuits?

[PartialDischarge]

Beacuse the dielectric strength varies from one axis to another, I’m talking under many tens of kVs, with pcbs operating in transformer oil, the limiting factor becomes not creepage or clearance but the paths inside the pcb through the fiber

[CZ101]

Great thread.

I have been investigating options for a high impedance differential preamp for looking at tiny signals in the audio band. My application right now is for probing differential microphone-level signals, somewhere between 5-50 mV RMS. The output would be single ended going into a typical scope 1 meg input.

The Tek AM502 would probably be exactly what I'm looking for but I'd kind of like to make this a DIY learning project and also see if lower noise is achievable.

I've been thinking about using the LSK489 as a differential JFET input pair:

http://www.linearsystems.com/lsnews/lsk489News.html
http://www.linearsystems.com/lsdata/datasheets/LSK489_LOW_NOISE,_LOW_CAPACITANCE_MONOLITHIC_DUAL_N-CHANNEL_JFET.pdf

Here's a nice app note from Bob Cordell on the LSK489:
http://www.cordellaudio.com/JFETs/LSK489appnote.pdf

The LSK489 would be in front of perhaps the AD8429 instrumentation amp http://www.analog.com/media/en/technical-documentation/data-sheets/AD8429.pdf. 1 Mhz bandwidth would be adequate, but higher would be welcome.

I'm right now trying to figure out the gain and probe matching options to optimize for CMRR and noise. The Tektronix probes with trimmable DC resistance (and capacitance) to match pairs for CMRR like the P6023 or P6135 are nice but they are 10x and it seems like these would really only be useful for larger signals, unless I am mistaken.

Then again the differential amplifier's rti noise decreases with gain, so maybe I am looking at this the wrong way and should go for 10x probes.

There is also the option of simply using a shielded twisted pair and making a diy probe that way.

[David Hess]

Beacuse the dielectric strength varies from one axis to another, I’m talking under many tens of kVs, with pcbs operating in transformer oil, the limiting factor becomes not creepage or clearance but the paths inside the pcb through the fiber

To prevent ionization where the dielectric constant falls?  I know this is an issue with bubbles in the potting material used in high voltage assemblies.

[David Hess]

I have been investigating options for a high impedance differential preamp for looking at tiny signals in the audio band. My application right now is for probing differential microphone-level signals, somewhere between 5-50 mV RMS. The output would be single ended going into a typical scope 1 meg input.

The Tek AM502 would probably be exactly what I'm looking for but I'd kind of like to make this a DIY learning project and also see if lower noise is achievable.

I've been thinking about using the LSK489 as a differential JFET input pair

At lower frequencies shunt feedback is more acceptable so you end up with a design like the Tektronix 5A22/7A22/AM502 using a differential JFET pair with at least enough gain to overcome the noise of the following low impedance stages.

If you do not want to mess around with the LSK489 and similar which will require considerable support circuitry, then parallel low noise JFET input operational amplifiers might be an acceptable simplification.

There is something to watch out for with dual JFETs however; this does not matter at audio frequencies but at higher frequencies, electrical coupling between the monolithic JFETs can ruin common mode rejection and transient response.  Jim Williams mentioned this in his own differential probe amplifier design shown below.  This may be why Tektronix used matched pairs in the 100MHz 7A13 instead of a dual part.  The slower 1MHz 5A22/7A22/AM502 used a dual part.

The Jim Williams design shown below is one of those with horrible overload recovery because of the slow DC common mode stabilizer loop.  It also is not optimized for low noise; there is no voltage gain in the JFET stage like with the 5A22/7A22/AM502.  Neither of these things mattered for his application.  Also note that he used one of those magical difference amplifiers on my list for differential to single ended conversion.

The LSK489 would be in front of perhaps the AD8429 instrumentation amp http://www.analog.com/media/en/technical-documentation/data-sheets/AD8429.pdf. 1 Mhz bandwidth would be adequate, but higher would be welcome.

Notice that there is a tradeoff with input noise, gain, and bandwidth for the AD8429.  If the JFET preamplifier has enough gain to overwhelm the noise, then the AD8429 can be operated at a lower gain resulting in increased bandwidth.

I'm right now trying to figure out the gain and probe matching options to optimize for CMRR and noise. The Tektronix probes with trimmable DC resistance (and capacitance) to match pairs for CMRR like the P6023 or P6135 are nice but they are 10x and it seems like these would really only be useful for larger signals, unless I am mistaken.

Then again the differential amplifier's rti noise decreases with gain, so maybe I am looking at this the wrong way and should go for 10x probes.

10x probes have two problems; the attenuation raises the input noise by the same amount (and more due to the 9 megohm series resistor but this is reduced at higher frequencies by the parallel capacitance) and the mismatch in attenuation lowers the common mode rejection a lot unless special matched probes are used.  The economical way to get around the matching issue is simply to add DC and AC trims to the 1 megohm shunt network at the preamplifier use a dedicated set of 10x probes.  The trimmable dual probes which Tektronix made did this with a network at the compensation box but had to pay for it by lowering the input series resistance from 9 megohms to a lower value.

There is also the option of simply using a shielded twisted pair and making a diy probe that way.

Or use 1x probes which will not degrade the input noise or common mode rejection.

Often you need some attenuation to get enough common mode input voltage range unless AC coupling is acceptable or your differential amplifier has a lot of input common mode range.  Unlike most oscilloscope inputs and bare active probes, the 5A22/7A22/AM502 and 7A13 have an input common mode range of +/-10 volts and the design we have been discussing with thanasisk also could have that much range just by using the right JFET operational amplifier and the AD830/AD8129/AD8130.

[Marco]

The LSK489 would be in front of perhaps the AD8429 instrumentation amp

Using the JFET only as an impedance transformer seems like a shame. There are circuits which preserve the noise level of the JFET.

[PartialDischarge]

Beacuse the dielectric strength varies from one axis to another, I’m talking under many tens of kVs, with pcbs operating in transformer oil, the limiting factor becomes not creepage or clearance but the paths inside the pcb through the fiber

To prevent ionization where the dielectric constant falls?  I know this is an issue with bubbles in the potting material used in high voltage assemblies.

Ionization and flashover. Sometimes ionization is not that much of a problem, like in equipments with low duty cycle, X-ray machines for human use generate up to 150kV and the duty is very low even in continuous fluoroscopy. However a flashover will make noise and sometimes render the equipment useless, maybe not by a technical fault per se, but as a reminder that high voltage has found its way around and will find it again.

Bubbles create a different problem, partial discharges. In oil there are no partial discharges since the electric field will shape a bubble, a droplet of water or a hair the way it wants to, to create flashover. In solid materials, like the epoxies used in instrument transformers, bushings or busbar insulators, a bubble or a small pebble inside represent a change in relative permittivities, which implies a redistribution of the field lines in that area. It's like having a capacitive divider inside the insulator. At some voltage, the inception voltage, the air in the bubble breaks down creating UV and ozone. With enough time the insulator develops cracks and fails. High voltage is like Tim Robbins in the Shawshank Redemption, it takes time but it will eventually escape.

High voltage 101 is thinking how the field and equipotential lines will behave. And they do according to the materials that they encounter. Sometimes a common mistake in high voltage systems is thinking that a thick piece of insulating material in between two air-insulated high voltage conductors improves reliability and that could be false, since insulating material with higher permittivity will drive away field lines to the outside, where now they become more condensed and breakdown could occur. Or place a low permittivity material inside a high valued one, and it will act as the small capacitor in a capacitive divider.

[David Hess]

The LSK489 would be in front of perhaps the AD8429 instrumentation amp

Using the JFET only as an impedance transformer seems like a shame. There are circuits which preserve the noise level of the JFET.

Both of those examples work the same way as the 5A22/7A22/AM502.  That is an awful lot of extra complexity for a 3dB noise improvement over JFET source followers driving a AD8429 configured for high gain.

I would like to see what kind of bandwidth that design driving a much higher noise AD8129  could produce.  I suspect it leads to all kinds of problems making JFET source followers desirable despite higher noise.

[thanasisk]

Have been quietly working on simulating a preamp based on LT1102 to have as a comparison reference against further designs that we have been discussing here,  however I do not seem to be able to achieve bandwidths of more than 50/550kHz (G of 100/10) despite the datasheet indication of approx 5 times that BW. Also the input and output noise is sky high.  Makes me wonder whether the pspice model from AD is ok or whether I am doing an obvious mistake in the design. Will post here my schematic hopefully late in the evening if time allows..

[David Hess]

When you post the schematic, also list the test conditions of the LT1102 circuit.

SPICE models do not always model every aspect of a part.  The better ones include a list of what is modeled.

[thanasisk]

Just throwing this in the pile, maybe it'll be useful to you at some point.

If you want to go all the way down to DC and up to say 10 or 20 Hz,  you can use a split path amp. You can have you ac-coupled amp,  say 10Hz to 1MHz  then,  also have something like a LTC1050  handling the DC part,  in parallel. Identical gain,  of course.

There's something  regarding this in App Note 106 fron  Linear.

There is also Application Note 47 from Linear:

http://www.analog.com/media/en/technical-documentation/application-notes/an47fa.pdf

They use a split path (Fig 74, 76 for gains of 10 and 1000 respectively). Interestingly an LT1102 handles the DC path, and a video difference high input impedance amplifier the AC path.

[thanasisk]

When you post the schematic, also list the test conditions of the LT1102 circuit.

SPICE models do not always model every aspect of a part.  The better ones include a list of what is modeled.

Here is the spice model. Not completely sure what is modelled and what is not:

http://www.analog.com/media/en/simulation-models/spice-models/LT1102.txt

I am attaching the schematic along with bode and noise plots (for several values of source impedance). I also attach the tina-ti/spice "analysis parameters".

I hope I am not doing an obvious stupid mistake with the schematic /design. G=20 as expected but BW much smaller than the data sheet. And noise is enormous.

[thanasisk]

Any ideas or feedback about the circuit? 

[Marco]

Properly DC bias the inputs, not through the diodes.

[David Hess]

Properly DC bias the inputs, not through the diodes.

Yep, I am not sure what SPICE does with unbiased JFET inputs but it probably is not good; either move the 1 megohm resistors to the amplifier side of the input capacitors for testing purposes or do what an oscilloscope input does and place like 100 kilohm resistors across the input capacitors.

1N4004 diodes are not suitable due to excessive leakage and capacitance.  People seem to like the BAV199 dual low leakage diode (3pA 2pF) but I like 2N3904 collector-base junctions.  The input capacitance of the protection diodes can be halved by using pairs in series for each one.

Note that the slow reverse recovery of the low leakage input protection diodes screws up overload recovery time.  Switching diodes like the 1N4148 are much faster but the gold doping which makes them fast also makes them have too much leakage for an application like this.  The base-emitter junction of a 2N3904 is both low leakage and fast but has a very limited reverse breakdown voltage.  They used to make 15 volt fast low leakage low capacitance diodes but there are other ways to do it if fast recovery of the input protection circuits is a requirement.

[thanasisk]

Well, Marco & David,  I placed the bias resistors after the input caps (see attached schematic) or even at the + and - inputs of LT1102 ; But this had no effect on the simulation results.

David I had in mind  the -GP version 1N4004GP which is Glass Passivated with 8pF junction capacitance, 5uA reverse current (at 100deg celcius at 400V peak reverse voltage)/(or just 20-50nA at 25deg celcius). Reverse recovery is 2usec though.

Why do you consider the 1N4148 as unsuitable? The worst case leakage current of 50uA is specified at 150 degrees celcius  BUT  at a more normal 25degrees it is in the region of 20-100nA depending on the (reverse) voltage.

Or do you follow worst case specs because you can expect a sudden junction temperature increase during overvoltage events?

As a side note, no reasonably priced through-hole part can come close to the specs the BAV199 that you mentioned.

[thanasisk]

And here is another take with an INA111. With similar layout as the previous posts for the LT1102. But it works this time! I will recheck the importing of the LT1102 model..

With the INA111 a gain of 2 brings the noise at the output above the noise floor of the AD2.

So I can now look for a suitable opamp with a gain of 10 and 100 to see what I can get. 

[David Hess]

David I had in mind  the -GP version 1N4004GP which is Glass Passivated with 8pF junction capacitance, 5uA reverse current (at 100deg celcius at 400V peak reverse voltage)/(or just 20-50nA at 25deg celcius). Reverse recovery is 2usec though.

But maximum reverse current at room temperature is 5 microamps and reverse current is only weakly related to reverse voltage.  The 8 picofarads of capacitance gets doubled with two diodes resulting in a total input capacitance high enough to present problems with probe compensation and bandwidth.

Why do you consider the 1N4148 as unsuitable? The worst case leakage current of 50uA is specified at 150 degrees celcius  BUT  at a more normal 25degrees it is in the region of 20-100nA depending on the (reverse) voltage.

20 nanoamps into 1 megohm is 20 millivolts and this leakage varies directly with temperature.

Worst case high temperature leakage current into the LT1102 inputs is about 2 nanoamps and leakage current at room temperature is more like 100 picoamps.

Or do you follow worst case specs because you can expect a sudden junction temperature increase during overvoltage events?

Leakage from junction temperature increases due to overload is actually a consideration for fast overload recovery.

It is worth mentioning where these relatively high absolute maximum leakage specifications come from.  It is expensive to test for low leakage (tester time is charged per second) so a specification like 5uA or 20nA represents what the tester itself was capable of within the time allowed for testing.  Where the LT1102 datasheet says +/-60pA maximum at 25C, +/-400pA maximum at 70C, and 15nA maximum at 125C, they are not kidding even though the typical specifications might be 10 times better.  2N3904 base-collector junctions are specified to be 50nA at 25C because of the test itself but are typically more like 10pA and better.

As a side note, no reasonably priced through-hole part can come close to the specs the BAV199 that you mentioned.

Dedicated low leakage diodes were never commonly available and technically the BAV199 does not count as low leakage either because it is only tested to be less than 5 nanoamps.

Manufacturers have used small signal transistor base-collector or base-emitter junctions as low leakage diodes for decades.  Like the BAV199, they are not tested for low leakage either but demand was never high enough to manufacture a tested low leakage part at an economical price.  The user has to qualify or test them themselves which is not difficult.

If you want an inexpensive already tested low leakage diode, then a 2N4117/2N4118/2N4119 low input bias current JFET (10pA at 25C maximum and 1pA typical) is the most economical choice and a lot of manufacturers use low leakage JFETs for exactly this purpose; just tie the drain and source together.

And here is another take with an INA111. With similar layout as the previous posts for the LT1102. But it works this time! I will recheck the importing of the LT1102 model..

That is weird; I was going to suggest that the SPICE was modeling the conductance of the 1N4004GP diode attenuating the input but obviously there is something wrong with the LT1102 model.  At zero volts, diode conductance is about 26 mhos/amp. (1)

(1) NIST's Guide for the Use of the International System of Units (SI) refers to the mho as an "unaccepted special name for an SI unit", and indicates that it should be strictly avoided. - Fuck you NIST.  I no longer accept any of your advice since you approved deliberately compromised NSA algorithms and standards for cryptography.  (2) You are not to be trusted with anything.  You are as bad as the FDA.  Die in a fire.

(2) Yea, this is just an excuse.  I would use mho in place of siemen anyway.  But NIST really is no longer to be trusted at least with anything having to do with cryptography and computer security.  They can still die in a fire.

[thanasisk]

David I had in mind  the -GP version 1N4004GP which is Glass Passivated with 8pF junction capacitance, 5uA reverse current (at 100deg celcius at 400V peak reverse voltage)/(or just 20-50nA at 25deg celcius). Reverse recovery is 2usec though.

But maximum reverse current at room temperature is 5 microamps and reverse current is only weakly related to reverse voltage.  The 8 picofarads of capacitance gets doubled with two diodes resulting in a total input capacitance high enough to present problems with probe compensation and bandwidth.

Ah yes, I see that the curves in the data sheet are "TYPICAL CHARACTERISTICS".

Why do you consider the 1N4148 as unsuitable? The worst case leakage current of 50uA is specified at 150 degrees celcius  BUT  at a more normal 25degrees it is in the region of 20-100nA depending on the (reverse) voltage.

20 nanoamps into 1 megohm is 20 millivolts and this leakage varies directly with temperature.

Worst case high temperature leakage current into the LT1102 inputs is about 2 nanoamps and leakage current at room temperature is more like 100 picoamps.

Or do you follow worst case specs because you can expect a sudden junction temperature increase during overvoltage events?

Leakage from junction temperature increases due to overload is actually a consideration for fast overload recovery.

It is worth mentioning where these relatively high absolute maximum leakage specifications come from.  It is expensive to test for low leakage (tester time is charged per second) so a specification like 5uA or 20nA represents what the tester itself was capable of within the time allowed for testing.  Where the LT1102 datasheet says +/-60pA maximum at 25C, +/-400pA maximum at 70C, and 15nA maximum at 125C, they are not kidding even though the typical specifications might be 10 times better.  2N3904 base-collector junctions are specified to be 50nA at 25C because of the test itself but are typically more like 10pA and better.

As a side note, no reasonably priced through-hole part can come close to the specs the BAV199 that you mentioned.

Dedicated low leakage diodes were never commonly available and technically the BAV199 does not count as low leakage either because it is only tested to be less than 5 nanoamps.

Manufacturers have used small signal transistor base-collector or base-emitter junctions as low leakage diodes for decades.  Like the BAV199, they are not tested for low leakage either but demand was never high enough to manufacture a tested low leakage part at an economical price.  The user has to qualify or test them themselves which is not difficult.

If you want an inexpensive already tested low leakage diode, then a 2N4117/2N4118/2N4119 low input bias current JFET (10pA at 25C maximum and 1pA typical) is the most economical choice and a lot of manufacturers use low leakage JFETs for exactly this purpose; just tie the drain and source together.

This is all very useful information, thank you!

All this discussion and using a jfet or transistor as diode replacement reminds me of the "biased diode clipping circuit" and the transistor clippers (e.g. BC456 and BC556 NPN/PNP transistors for diode replacement) analyzed in Chapter 24 of Small Signal Audio Design by Douglas Self. There he also mentions that circuits like that have a very high output impedance and thus he always includes an followup op-amp buffer. I wonder whether such a concern is relevant to our case here.

I will certainly try to test the JFET solution though, any pointers to relevant schematics of manufacturers?

And here is another take with an INA111. With similar layout as the previous posts for the LT1102. But it works this time! I will recheck the importing of the LT1102 model..

That is weird; I was going to suggest that the SPICE was modeling the conductance of the 1N4004GP diode attenuating the input but obviously there is something wrong with the LT1102 model.  At zero volts, diode conductance is about 26 mhos/amp. (1)

(1) NIST's Guide for the Use of the International System of Units (SI) refers to the mho as an "unaccepted special name for an SI unit", and indicates that it should be strictly avoided. - Fuck you NIST.  I no longer accept any of your advice since you approved deliberately compromised NSA algorithms and standards for cryptography.  (2) You are not to be trusted with anything.  You are as bad as the FDA.  Die in a fire.

(2) Yea, this is just an excuse.  I would use mho in place of siemen anyway.  But NIST really is no longer to be trusted at least with anything having to do with cryptography and computer security.  They can still die in a fire.

Well as long as you pass on unambiguous info, any term should be fine

I further tested an opamp post-amp configuration and will post my findings here soon. I got stuck a bit because of the compensation and because I also want to simulate for adequate THD+N performance..

[David Hess]

If you want an inexpensive already tested low leakage diode, then a 2N4117/2N4118/2N4119 low input bias current JFET (10pA at 25C maximum and 1pA typical) is the most economical choice and a lot of manufacturers use low leakage JFETs for exactly this purpose; just tie the drain and source together.

All this discussion and using a jfet or transistor as diode replacement reminds me of the "biased diode clipping circuit" and the transistor clippers (e.g. BC456 and BC556 NPN/PNP transistors for diode replacement) analyzed in Chapter 24 of Small Signal Audio Design by Douglas Self. There he also mentions that circuits like that have a very high output impedance and thus he always includes an followup op-amp buffer. I wonder whether such a concern is relevant to our case here.

The use of the transistors instead of diodes in Audio Design (great book, recommended) was because of higher conductance producing a slightly sharper knee in the clamp circuit.  Audio circuits have much different requirements than oscilloscope inputs because they can tolerate much less distortion.

Diode conductance (26 mhos/amp at zero volts) is a real thing and in high impedance circuits it can cause real problems if low leakage diodes are not used.  For oscilloscopes, up to a certain point capacitance is more important given the frequency requirements.  You *could* get away with gold doped high leakage 1N4148s but their leakage would cause considerable DC drift over temperature at 1mV/div in a 1 megohm high input impedance amplifier or buffer.

Oscilloscope front ends are not particularly linear and suffer from considerable distortion in the quest for high bandwidth and good transient response.  A front end intended for low distortion measurements not far from the audio band requires a different circuit topology including different input protection circuits.  For low distortion where input protection was still required, I would look to bootstrapping the protection circuits themselves so the clamp voltages follow the input voltage up to the clamp limits.  This is common in high performance multimeters and even some oscilloscopes do this.

I will certainly try to test the JFET solution though, any pointers to relevant schematics of manufacturers?

I am not sure what you are asking.  The 2N4117/2N4118/2N4119 JFETs (and the A versions) have been the lowest cost guaranteed low leak current diodes for a long time.  Other JFETs will work as well but most are not specifically tested for low leakage unless they are intended for applications which require low leakage.