Looking for a low noise amplifier (<1nV / rt. Hz)

[jbb]

I’m working on a wide band current sense system (DC - 1MHz) and I’m hunting for a better amplifier.

I checked Analog Devices (and this LT) and Texas Instruments and came up with the ADA4895-1, which will run off my +- 3.3V analog rails and has about 1nV/rt.Hz ven.

The LT1028 is a tiny bit better but consumes a lot more current and needs +-5V rails.

Can anyone suggest another supplier which might have something even better?

[coppercone2]

I don't think you will get much better then those without parallel devices.

[ejeffrey]

I have used the OPA211/OPA2211 which is 1.1 nV/rtHz and would work on +/- 3.3 V. 

At this level, remember that the noise density is usually specified at 1 kHz.  If you want to compare 0.9 nV, 1.0 nV and 1.1 nV, you really need to look at the performance at the frequency you care about.  This means including 1/f noise at low frequency and including the effect of input capacitance on noise gain at high frequency.

Its really hard to do much better than 1 nV/rtHz, especially at low frequency.  Some HEMT amplifiers do a factor of 2-3 better than that but only at high frequency, pretty much anything beyond that requires cryogenic cooling.  Most current shunts are terribly matched to transistor amplifiers which have noise resistances in the hundreds or thousands of ohms.  If you could live with AC coupling a transformer would really help.  As an illustration, look at the SR554 which gets a noise voltage of 0.1 nV/rtHz by using a 100:1 transformer.

[SiliconWizard]

Maybe the LT6200: 0.95nV/√Hz (100kHz)
A lot more power hungry.

[jbb]

Thanks all

I had a look at paralleling 2 op-amps, and it made some improvement (approx 20% drop in total system noise). I don’t think that’s worth it on my 1st prototype where I will likely have some other mistakes to fix.

Thank you for mentioning the SR554, ejeffrey. Right now I’m looking for DC - 200 kHz (or 1 MHz) bandwidth, so a straight up transformer input won’t hit the spot. However, in version 2 or 3 I could have a look at a LF + HF path to get a wide band transformer in there.

[Marco]

A discrete differential pair front end with say 9 BJTs on each leg?

[drussell]

Does Linear Systems make anything suitable to adapt a circuit around?  I know they make all sorts of very low noise stuff, certainly lots of JFETs and I remember seeing some article in some print journal that I'm sure I still have around here somewhere years ago about using some LSK parts for an op-amp front end to get a low noise configuration but I think they were mostly going for very high input impedance rather than ultra low noise.

I honestly don't recall.   Hmmm...  Think think think... 

[Gerhard_dk4xp]

I've pushed it to 220 pV/rt Hz with 10 pairs of ada4898-2, but they also need +-5V.
The 4898-2 is interesting because you get 2 op amps for the price of an AD797.
<   http://www.hoffmann-hochfrequenz.de/downloads/lono.pdf   >
My preamp has been redesigned in the mean time with some costly components :-(
such as RS-170 relays and a wet tantalum cap in the input.

There is also a differential preamp in Art of Electronics, ed 3.  page 506.. I first wanted to build
the differential version, but since I needed that ugly coupling cap anyway, I settled on
the single ended version. It features 70 pV/rt Hz and needs only 1/4th the transistors.

regards,
Gerhard

[Marco]

Need the differential (ie. composite opamp) for DC since he doesn't want a two path amplifier.

[jbb]

Need the differential (ie. composite opamp) for DC since he doesn't want a two path amplifier.

Yes, I'm looking at composite amplifier for (at least) the first go.  Future versions - if I decide to keep going - can go further afield.

So I had a bit of a poke about online and found this beast from the good people at ETH Zurich.  100nV/day drift, 100fA max input leakage, Peltier cooling and an oven.  Quote of the paper:

... It is therefore essential to test the components prior to assembly: the input FET, some active components and the leakage current of some capacitors. Only a fraction of the components fulfils the requirements and can be used...

Decoded: "we spent ####ing ages binning components and most of them weren't good enough."

[cat87]

I've come across some designs until now that might be what you need. Bust they don't come in a nice SO package.

There's this amp using the 2SK117,  claims 1.4 nv/sqrtHz @ 1kHz:
http://techlib.com/electronics/audioamps.html

Then there's Glenn's design:
http://www.glensstuff.com/ulnma/ulnma.htm

And the eevblog discussion thread on this one is:
https://www.eevblog.com/forum/projects/low-noise-amplifier/

For Glenn's design,  Rs still has some of those BFxxx Jfets in stock.

Hope these help.


Later edit:
Just remembered there's also a nice design in LT's application note 159. The design there is mainly for low frequency stuff but it can  go up to 1Mhz and more
 

[Gerhard_dk4xp]

Yes, I'm looking at composite amplifier for (at least) the first go.  Future versions - if I decide to keep going - can go further afield.

... It is therefore essential to test the components prior to assembly: the input FET, some active components and the leakage current of some capacitors. Only a fraction of the components fulfils the requirements and can be used...

Decoded: "we spent ####ing ages binning components and most of them weren't good enough."

Exactly.  These here are the Id vs. Vgs plots for the IF3602. Don't expect them to be matched in the sense
we use that word. But at least you don't get opposite outliers in one can for €55 per piece. (Mouser)

<    https://www.flickr.com/photos/137684711@N07/37321004540/in/album-72157662535945536/lightbox/   >

You also get a few hundred pF input capacitance per transistor, so cascoding is definitely required for your MHz.
The ADA4898 does not share that problem. 1 MHz is easy.
In that FLICKR album are also measurements of some depletion FETs that can be used as a cascode without
gate voltage dividers. That eases bootstrapping the JFET's drain if required.

On Semi has some new JFETs that can replace the EOLed BF862.

I have not yet read this ETH article, but in JFETs Id and Vgs have opposite TCs, so there is an
operating point that has TC = 0, and one may get by with less oven ado.

If you don't want to count single electrons, you will probably have a low impedance shunt resistor,
so better go bipolar.   (A sub-nV/rtHz preamp fed from a 1K source is futile a priori.)

In the Mat-02, 03, 04 data sheets are also examples for multiple transistor pairs in front of
op27-ish op amps, but parallel op amps are probably more painless.

cheers, Gerhard

[Marco]

If you can use a wideband transformer, you don't need FETs ... your source impedance will be very low. If your source impedance is very low you can use BJTs.

[Marco]

BTW, has anyone ever made a low noise amplifier out of boost converter? I've never seen it, but it should be possible.

[Echo88]

@Marco: Wouldnt the input noise be massively changed trough the inductor?
@Gerhard: Why do you use wet tantalums now? Too much leakage current/leakage current noise from the foil caps that you use?
Thanks for the paper jbb, very interesting!

[Gerhard_dk4xp]

That's because I underestimated the need for a large input capacitor.

A 100 or 200 uF foil capacitor would have been enough from a frequency
response point of view, but that resulted in a rise of noise at the bottom end
that is much stronger than 1/f.

Sub-nV/rtHz amplifiers make only sense when their input signal source is
smaller than, say, 60 Ohms. 60 Ohms delivers 1nV/rt Hz thermal noise at
room temperature and your complete system could never be better than that.

OTOH the input impedance of the preamp itself is high and determined
mostly by the 10K bias resistor. But these 10K also create noise like a 10K
resistor. That completely spoils the good noise behavior of the preamp itself.

Here the low signal source impedance can help. AC-wise, it is in parallel to
the bias resistor, but the coupling C is in series to it. That works nicely in the
middle of the passband of the amplifier.

But at the lower -3dB point, Xc is equal to Rbias, so there is not much shorting
left. From a noise viewpoint, Xc must be small against the signal source impedance
and not only small against Rbias.

That is very unpleasant. You no longer need 100 uF but 5000 uF.
That cannot be made with foil caps. I already had a dozen or two of WIMA
10u/50V in parallel. The largest foil caps I could get and already costly.

So i used the wet slug tantals. 5000u cost about €70 if you are very lucky.
Could be more than twice that. The cheapest I have found was AVX from Digikey.
Vishay was MUCH more than twice.

The wet tantalums still have a guaranteed ESR of 400 mOhm. Quite a lot, IMHO.
I choose the wet slug tantal after reading Jim Williams' article on measuring the ref noise
of the LT6655.

BUT, even in the 70 pV amplifier in the Art-of-Electronics-III-style i could not find
an obvious drawback using good quality aluminium electrolytics by Panasonic. YMMV.
That design was more a proof-of-concept for a chopper amplifier. At the 500 KHz
chopping frequency, the capacitors are harmless.

In FET amplifiers you may see gate resistors in the 100 Meg range and one is
tempted to see that as a drawback, noise voltage-wise. But that's not true.
The 100 Meg are easily ac-shorted.


Pic: chopper amp, might feature 100 pV rt Hz at << 100 mHz, but I cannot yet verify
that because of the 1/f noise of my 89411A FFT analyzer. The CPLD creates the
chop clock and the delayed unchop-clock after the 500 KHz filter.

[Echo88]

Interesting details, thanks. Didnt know you design discrete chopper amps yourself. Did you consider using two LNAs and cross-correlation to further reduce your LNA-noise? Like the battery noise measurement at NIST for example.
I have a cookie-box with two 0.1-10Hz-LNAs and would like to employ the cross-correlation via Python, but lack the knowledge/documents to do so.

[Gerhard_dk4xp]

I have played shortly with the cross correlation in my Agilent 89441A.
The effect was so strong that I had doubts about the results. That
machine has so many things to set up that you need a program
for it (or at least a checklist).

I have just married a LT2500-32 ADC to a Beaglebone Black; that
has probably much better 1/f behavior than the 89441A. I'm just
getting first time series; still a lot of software to do. Eventually I'll do
cross correlation also, maybe over xmas.

It seems that numpy can do cross correlation, but my programs
are in C.

cheers, Gerhard

[David Hess]

A discrete differential pair front end with say 9 BJTs on each leg?

It does not even require parallel devices to do this.  Large area audio driver transistors like the BC327/BC337 can operate below 1nV/SqrtHz assuming that the relatively high bias current is acceptable.  A precision amplifier is used in parallel to remove DC offset.  If low noise at low frequencies is necessary, then a chopper stabilized amplifier will be needed to remove 1/f noise.

The BC327/BC337 is not used for its power or current handling capability.  The large area is a way to lower the base spreading resistance which limits low noise operation.  Using lots of smaller transistors in parallel is more complicated because emitter ballasting cannot be used without raising the noise level.  This points to the problem JFETs have achieving low noise with their lower transconductance from their higher source resistance.