Ultra Low Noise JFET

[TimFox]

Asking from a point of complete ignorance, how well could vacuum tube circuits do in comparison to low noise semiconductor circuits? I'm thinking about the old Tektronix 1A7 high gain diff plug-in that used parallel Nuvistors. Were they remotely competitive to what we can do today?

Tubes still have shot noise, and other noise sources like from secondary emission.

I remember that different tube types, triode, tetrode, and pentode, had different noise characteristics.

I was surprised that common tubes have relatively high grid current compared to the gate leakage current of JFETs.  Only specialty "electrometer" tubes had low grid current, so JFETs (and super-beta bipolar transistors) were a big improvement.

I would expect that the dependence of valves on heaters would make them moderately noisy, but they should avoid the 1/f noise of semiconductors, unless they have a similar mechanism.

1/f noise can be very similar to thermal drift, so it would not surprise me if the high operating temperatures interacting with radiative cooling produce the same effect.  In precision solid state circuits, it can be difficult to distinguish 1/f noise from low frequency thermally induced drift.

Tube noise: 
The theoretical input noise voltage density for a small-signal triode operated in normal (negative grid) circuits, space-charge limited, is usually stated (in older books) as that of an equivalent noise resistance approximately equal to
R_N = 2.5 / g_m .
Adding extra grids to make a pentode increases that because of the statistical fluctuation of how many electrons emitted by the cathode flow to the screen instead of the plate ("partition noise").

Grid current is an interesting phenomenon:  at DC and low frequencies, it varies with grid-cathode vooltage.  At RF frequencies, there are other mechanisms that increase the noise current due to coupling from the cathode emission.
It is often ignored when the grid is negative with respect to the cathode, although it is macroscopic with the grid positive, forming a diode with the cathode.
However, there are several sources for grid current, which have different polarities.  It is possible to find a range of negative grid-cathode voltages where the different components cancel, giving a net current that is very low.
Attached is a graph of grid current, along with plate and screen currents, that I measured for a high-gain pentode (E180F).

Although this shows a region of very low grid current, since it is a sum of different paths the standard deviation (noise) of that current can be important.
In general, one should not run such a tube at very low grid bias (highest cathode current), since low grid reverse bias does not stop the energetic electrons emitted by a hot cathode.
Another reason to avoid grid current is that away from the "plateau", the grid current is a non-linear function of grid voltage and can induce distortion when driven from a high impedance source, such as the output impedance of the first stage.  (Guess if I learned that the easy or hard way.)

[G0HZU]

To try and demonstrate a bit of confidence in the setup, I changed the bias point of the ZTX851 to 1mA and the Genesys simulation predicts the noise figure will degrade to about 1.1 dB at a few kHz. See the simulation plot below.

I then re-biased the jig to 1 mA Ic and pressed the GET NF button after it had stabilised, and got a noise figure just under 1.2 dB at a few kHz for the 'real' ZTX851 in my text fixture

[G0HZU]

I can't promise anything, but I'll try and measure some JFETs tomorrow.

I've got classic process 50 JFETs like the 2N4416A and BF256B here. Also process 92 devices like U310 and J310. I also have CPH3910 and various chopper types from other processes. I don't have any really exotic types here but I do have genuine NOS U320 JFETs here still in the original ESD packaging from new. These would normally be used up at RF as would most of the other JFETs I've listed. I've got lots of experience using these parts in RF amplifiers up at HF and VHF but very little using them at audio except with Hi Z microphones.

Which devices are of interest and what operating point and source impedances are of interest?

[ArgyllGargoyle]

I’m interested in cph3910 at low source impedance- I heard it was the best low noise replacement for bf862

[Gerhard_dk4xp]

I got about 320 pV/rtHz from 16 pcs.  cph3910.

For noise in FETs discussions see Felix Levinzon's book

© Springer International Publishing Switzerland 2015
F. Levinzon, Piezoelectric Accelerometers with Integral Electronics,
DOI 10.1007/978-3-319-08078-9_5

I think I heard someone mumble that chapter 5 is on sci-hub, along with
many pointers to interesting papers.

If your 1/f noise is not 1/f but MUCH steeper, then probably your input
coupling capacitor is too small. The noise of your bias network then is not
properly shorted. I had that in my 10 * 2 *ADA4898 preamp
(abt. 12*10uF foil). Scott Wurzer from AD (AD825, AD797...) saw
that on the spot.
That has nothing to do with the lower f-3dB. A much bigger cap
is needed.

I have built the transistor noise measurement setup from Art Of Electronics
ed 3. Works as promised. Also their 70 pV/rt Hz ribbon preamp.
I made a non-differential version, also OK.

You don't need a signal generator for calibration. A 61Ohm resistor
at room temp. delivers a convenient calibration line at 1 nV/rt Hz
directly from first principles. And that without noise pickup loops
if cold switched and on the board.

Gerhard

[Gerhard_dk4xp]

That new safety feature is a royal pain. You modify your text,
get checked again and then your text has returned to an even
older version.

[moffy]

@Gerhard_dk4xp, curiosity about the eight relays on the first photo, what are they for?

[moffy]

Here is an interesting article about the 1/f noise of electrolytic capacitors and how it can be used as a reliability indicator.
https://www.diyaudio.com/community/attachments/1-over-f-noise-of-electrolytic-capacitors-as-a-reliability-indicator-pdf.1262206/

[moffy]

I can't promise anything, but I'll try and measure some JFETs tomorrow.

I've got classic process 50 JFETs like the 2N4416A and BF256B here. Also process 92 devices like U310 and J310. I also have CPH3910 and various chopper types from other processes. I don't have any really exotic types here but I do have genuine NOS U320 JFETs here still in the original ESD packaging from new. These would normally be used up at RF as would most of the other JFETs I've listed. I've got lots of experience using these parts in RF amplifiers up at HF and VHF but very little using them at audio except with Hi Z microphones.

Which devices are of interest and what operating point and source impedances are of interest?

Any devices you test will be of interest, though personally I am interested in the low frequency noise performance. But that said, having the data from a number of devices is very useful.

[G0HZU]

I don't think the caps are currently the dominant issue with my test fixture, I think it's the active filter on the Vcc line. I know that if I go back to using a battery here, then it cleans up nicely below 50 Hz when testing the ZTX851 and there is just a small amount of flicker noise, so small it isn't easy to see unless I use lots of correlations. There's only 0.6V DC at the input cap so I think it's less likely to generate noise with such a low bias. Note that I've also measured the tant caps for ESR right down to 10 Hz and they seem fine in terms of how much extra resistance they add inline.

With a JFET I might see several volts across the input cap so things might be different then. I won't know until I try. This stuff can be incredibly time consuming and I have lots of other stuff to do today. But I'll try and measure a couple of JFETs with something like a 470R source impedance and an Id of 5-8 mA or so. I'll use a battery for the main supply.

[moffy]

I don't think the caps are currently the dominant issue with my test fixture, I think it's the active filter on the Vcc line. I know that if I go back to using a battery here, then it cleans up nicely below 50 Hz when testing the ZTX851 and there is just a small amount of flicker noise, so small it isn't easy to see unless I use lots of correlations. There's only 0.6V DC at the input cap so I think it's less likely to generate noise with such a low bias. Note that I've also measured the tant caps for ESR right down to 10 Hz and they seem fine in terms of how much extra resistance they add inline.

With a JFET I might see several volts across the input cap so things might be different then. I won't know until I try. This stuff can be incredibly time consuming and I have lots of other stuff to do today. But I'll try and measure a couple of JFETs with something like a 470R source impedance and an Id of 5-8 mA or so. I'll use a battery for the main supply.

I agree, I don't think the tant cap is contributing any significant noise to your jig. The correlation between transistor datasheet and your results say that. I put the reference there just for general info and something to be aware of. We appreciate your time and effort and understand how involved this can get. Take your time.

[G0HZU]

OK thanks...
I have just tried putting a U310 (process 92) JFET into the fixture with caps added/changed to suit a negative gate bias. I've had to select the 1k drain resistor as I'm testing at about 6mA Id and I need a sensible Vd voltage. The transconductance of these JFETs is low compared to a BJT at this bias current and so the secondary noise contribution from the 1 k resistor and the LT1677 will add a tiny bit to the measured noise figure.

However, with a 470R source resistance and 6 mA Id, the U310 model in Genesys predicts a noise figure of about 0.9 dB at 10 kHz when simulated. I haven't studied the U310 model but it does show a gradual degradation in noise figure below a few kHz. So I was expecting to see relatively poor performance below 1 kHz with the real U310 in the test jig.

When I measure the real U310 with similar test conditions, I measure a noise figure of 1.1 dB but I have to measure this above 3 kHz because the flicker corner frequency appears to be about 3 kHz. So this part looks to be a poor choice under these test conditions as the noise figure rapidly degrades below about 1 kHz. So this device doesn't seem to be a good choice here.

I think I need to spend some time thinking about how to make a jig optimised for JFET testing. This will probably mean using a different opamp than the LT1677 and a different drain load resistance. I think my current jig is just about OK but I will need to do more testing. I won't be able to make a dedicated JFET jig for quite a while. I'd need to order some parts and make a PCB etc.

[zike]

Tubes still have shot noise, and other noise sources like from secondary emission.

Tube noise: 
The theoretical input noise voltage density for a small-signal triode operated in normal (negative grid) circuits, space-charge limited, is usually stated (in older books) as that of an equivalent noise resistance approximately equal to
R_N = 2.5 / g_m .
Adding extra grids to make a pentode increases that because of the statistical fluctuation of how many electrons emitted by the cathode flow to the screen instead of the plate ("partition noise").

How does this thermal component compete with shot noise in real applications? I think one expects \$ e_n = I_n / g_m = {\sqrt{2eI_{DC}}}/g_m \$  referred to input. For, say, a 12AX7 at 1.2 mA, \$g_m \approx 1,500 \mu S \$ this comes to about 13 nV/\$\sqrt{\rm Hz} \$ which would exceed the above (about 5 nV/\$\sqrt{\rm Hz} \$ for this \$g_m\$ at room temperature). 

Which leads me to also ask, what's the "temperature" of this noise-equivalent resistor?

On another tangent I've restored some UHV instruments with CK5889 and 5886 electrometer tubes, these ran at very low plate voltage (~ 12V) to achieve ~ femtoamp grid currents. 

[G0HZU]

I made up a little SMD adaptor PCB (with legs) to suit a SOT-23 packaged CPH3910 JFET and just put one into the test jig. I didn't spend much time on this and set the Vgs for just under 5 mA Id. The source resistance was kept at 470R.

With this JFET, the flicker corner is now much lower at about 250 Hz and I measure a very repeatable noise figure of 0.50 dB at 1.2 kHz. I don't currently have a model for this JFET that I can use in Genesys but I'll see if I can find one later.

[TimFox]

Tubes still have shot noise, and other noise sources like from secondary emission.

Tube noise: 
The theoretical input noise voltage density for a small-signal triode operated in normal (negative grid) circuits, space-charge limited, is usually stated (in older books) as that of an equivalent noise resistance approximately equal to
R_N = 2.5 / g_m .
Adding extra grids to make a pentode increases that because of the statistical fluctuation of how many electrons emitted by the cathode flow to the screen instead of the plate ("partition noise").

How does this thermal component compete with shot noise in real applications? I think one expects \$ e_n = I_n / g_m = {\sqrt{2eI_{DC}}}/g_m \$  referred to input. For, say, a 12AX7 at 1.2 mA, \$g_m \approx 1,500 \mu S \$ this comes to about 13 nV/\$\sqrt{\rm Hz} \$ which would exceed the above (about 5 nV/\$\sqrt{\rm Hz} \$ for this \$g_m\$ at room temperature). 

Which leads me to also ask, what's the "temperature" of this noise-equivalent resistor?

On another tangent I've restored some UHV instruments with CK5889 and 5886 electrometer tubes, these ran at very low plate voltage (~ 12V) to achieve ~ femtoamp grid currents.

Details:
In a vacuum tube operating under “space-charge limited” conditions, the shot noise is modified by space-charge effects.  The shot noise formula is valid under “saturated emission” conditions, where the plate voltage is high enough to grab every electron emitted from the hot cathode.  The 5722 noise diode was designed to work that way for measuring noise figure.  In full space-charge limit, the plate current is independent of cathode temperature, but in saturated emission it is a strong function of cathode temperature.  A 12AX7 is always operated in space-charge limit.  The resistor in the equation I quoted is at room temperature; it gives a simpler formula than that for noise voltage density itself.  When that model works, one compares the equivalent resistance to the actual source resistance to see how much noise the tube adds to the unavoidable noise from the source.
Electrometer tubes are very interesting: they have lousy gain but achieve low grid current by minimizing photo electron emission (due to low-energy x rays from the plate), among other optimizations.
Excess (1/f) noise in tubes is hard to predict, usually blamed on defects.

[G0HZU]

To help to minimise uncertainty, the 20dB attenuator I use is a precision device I designed many years ago as a calibration aid up into UHF. I designed it to achieve ultra low VSWR and ultra accurate 20 dB attenuation accuracy. I've posted up the plots for it before, but here they are again. The attenuator should typically be within +/- 0.02dB over a wide frequency range and I've measured it various ways, either with a VNA or various DMMs and analysers. To get the accuracy, each resistance is made up from two resistors hand selected using a 6.5 digit 4 wire ohmmeter and it was designed and optimised using a computer. A real labour of love. The divide by 500 attenuator should have similar accuracy as I made this with selected parts too.

The scale fidelity of the E5052A seems to be really good. I have various measuring receivers here that have excellent scale fidelity over a huge log range and the E5052A seems to be similar. The fidelity plot below is for one of my measuring receivers. I think the other one is slightly better than this and I'd like to think the E5052A will be in this ballpark at least. So I'd like to think the overall uncertainty of my noise figure measurements is quite low.

With this in mind, I've downloaded the manufacturer's model of the CPH3910 and put it into Genesys and then played with the operating point and source impedance to try and find the lowest noise figure at a few kHz. It falls below 0.2dB which is very impressive (if true) and I tried testing with this source impedance (about 2 k ohm) and an operating point of 5 mA Id. The noise figure is a bit lower up towards 10 mA on the simulator but it's difficult to arrange this with the jig in its current state. So I tested at 5 mA Id.

I measured it about 20 times in quick succession and got anything from 0.05dB through to 0.25dB for the noise figure but it was mostly 0.16 dB which was very close to the simulation. This is encouraging for my test setup. When I set out to do this noise figure testing with the E5052A I was hoping for < +/- 0.5 dB uncertainty but really wasn't sure what the E5052A could deliver. The Keithley 2015 THD helps here too as it has excellent accuracy for AC voltage measurements at a few kHz. So I think this all explains why I'm getting good results so far.

Over time, I may find some errors creeping in due to thermal drift in the test gear and I also have to be mindful of the room temperature when measuring noise figures this low. I have to update the spreadsheet with the correct ambient temperature near the test jig.

I've included a quick Visio diagram showing the test setup for testing a BJT. This shows the Vcc being powered from the low noise port of the E5052A. However, even this causes a tiny amount of excess noise below 50 Hz. So for critical stuff I have to use a battery here. I think I can improve the active filter on the test jig (so I don't need the battery) but I probably won't bother, at least for a while.

[G0HZU]

One feature of the Genesys model is that it allows a flicker noise coefficient to be entered. Genesys is very powerful and this coefficient can be tweaked fluidly in real time whilst playing with the source impedance and operating point as scrollable variables. So I first tried to get the best fit to the measured flicker noise of the real CPH3910 with a 470R source impedance. Once this was found and locked as a fixed flicker noise coefficient, I then played with the operating point and source impedance and the simulator suggests that there is an operating point and (fairly low) source impedance for maintaining very low noise figure (still sub 1 dB) and very little flicker noise down to 10 Hz. Maybe a 1dB rise down at 10 Hz. I tried it and it gave very similar results with the real hardware with hardly any excess noise visible. So the model seems to be quite useful. I now need lots more time to get used to playing with JFETs down at AF and I also need to improve the test jig quite a bit I think.

Of course, I'd expect there to be quite a spread in flicker noise between samples of real CPH3910 JFETs so I'm not sure how repeatable any of this stuff will be from device to device.

[EC8010]

@TimFox: That grid current in post #50 looks staggeringly high. My grid current measurements for small-signal valves tend to be single digit nA until the grid starts attracting electrons.

Back to the topic of JFETs and noise. I've not been impressed by InterFet devices. But Toshiba 2SK2145BL (available at a sensible price from Mouser) is quite quiet, and it's certainly quiet if you parallel a few. I've made <1nV/root Hz LNAs with them. I measure their noise by configuring the DUT as a differential pair driving an NE5532A and setting closed loop gain of 5000, low-pass (LC) filtering the output to 1kHz or less (to prevent oscilloscope overload), then taking the noise to an oscilloscope and averaging across 100 FFTs. I use lead-acid battery power and biscuit tins to avoid hum problems. My noise measurements generally confirm manufacturers' data, although I often go lower in frequency, necessitating very long measurement time (28 hours).

[Gerhard_dk4xp]

@Gerhard_dk4xp, curiosity about the eight relays on the first photo, what are they for?

bottom right = Axicom relay = speed up bias loop until close to O/P
                    ( would take time until the cows come home otherwise)

bottom left = use 1nV/rtHz from 61 Ohms or main input

above left = select direct input or FET stage for post amplifier

one in the top row: Bandwidth limit to 10 KHz

others:  40 to 80 dB dB gain in 10 dB steps + 3 dB gain boost   (from main input)

I had a version with analog switches instead of the relays, but signal
isolation was less than ideal. My soldering iron and the white relays
seem to have a love/hate relationship.

[TimFox]

@TimFox: That grid current in post #50 looks staggeringly high. My grid current measurements for small-signal valves tend to be single digit nA until the grid starts attracting electrons.

Grid current:  the E180F is a frame-grid tube with close grid-cathode spacing.  Checking my earlier measurements, I see that this case had a relatively low screen-grid voltage.
The graph covers a wide range of grid-cathode voltage.
I looked back at the original .xlsx file, and the grid current is only a few nA over a reasonable range of operating conditions (with useful transconductance), but that is obscured in the flat region on my graph.
I couldn't use log scales, since the current changes polarity as the grid voltage approaches zero.

Here's a summary of data taken with a range of tubes, triodes and pentodes.
I wanted to compare them at a normal operating cathode current value.

 GridCurrent2.pdf (85.77 kB - downloaded 20 times.)

[moffy]

With this in mind, I've downloaded the manufacturer's model of the CPH3910 and put it into Genesys and then played with the operating point and source impedance to try and find the lowest noise figure at a few kHz. It falls below 0.2dB which is very impressive (if true) and I tried testing with this source impedance (about 2 k ohm) and an operating point of 5 mA Id. The noise figure is a bit lower up towards 10 mA on the simulator but it's difficult to arrange this with the jig in its current state. So I tested at 5 mA Id.

I have never heard of the Genesys modelling before, but being able to modify parameters in real time is very impressive. But of course it requires the infrastructure of your testing. For the few low noise measurements I've made (very few)  I have  +/-12V batteries which I've found essential for getting rid of the 50Hz harmonics in my limited setup.

[Conrad Hoffman]

So, why did I ask the silly question about tube noise levels? Long ago I was involved in the design and construction of atomic force microscopes. Binnig, Quate and Gerber in speculated about atomic force imaging in 1986. Actual images came a bit later. It turns out these things are not that difficult to build if you know some of the tricks for stability. I've always found it an interesting thought experiment to ponder how long ago you could have built one. The scanning is accomplished with piezoelectric elements, but we've had Rochelle salt since 1675. Or use quartz. Once tube amplifiers came along, an oscillator and scanning system could be built. A CRT with persistence could display the results, or even one of those meter movements where a stylus draws on a piece of smoked glass. IMO, the hard part is you need a very low noise amplifier. We used to use OPA627 opamps in commercial designs. My guess is it could have been done with a carefully designed tube amp. Maybe. I don't know what noise level would be needed. Anyway, I speculate that an atomic force microscope could have been built in the 1930s, maybe even the 1920s in very crude form.

[TimFox]

So, why did I ask the silly question about tube noise levels? Long ago I was involved in the design and construction of atomic force microscopes. Binnig, Quate and Gerber in speculated about atomic force imaging in 1986. Actual images came a bit later. It turns out these things are not that difficult to build if you know some of the tricks for stability. I've always found it an interesting thought experiment to ponder how long ago you could have built one. The scanning is accomplished with piezoelectric elements, but we've had Rochelle salt since 1675. Or use quartz. Once tube amplifiers came along, an oscillator and scanning system could be built. A CRT with persistence could display the results, or even one of those meter movements where a stylus draws on a piece of smoked glass. IMO, the hard part is you need a very low noise amplifier. We used to use OPA627 opamps in commercial designs. My guess is it could have been done with a carefully designed tube amp. Maybe. I don't know what noise level would be needed. Anyway, I speculate that an atomic force microscope could have been built in the 1930s, maybe even the 1920s in very crude form.

At a high-end EKG manufacturer's headquarters, I saw a museum display of EKG equipment going back to before vacuum tube amplifiers, where quartz-fiber devices were used to write on smoked glass to record waveforms. 
Then came vacuum-tube electronics and pen-chart recorders.

[EC8010]

I couldn't use log scales, since the current changes polarity as the grid voltage approaches zero.

Here's a summary of data taken with a range of tubes, triodes and pentodes.
I wanted to compare them at a normal operating cathode current value.

(Attachment Link)

Yes, it's a nuisance that grid current has a wide range and changes polarity. Your spreadsheet numbers are much more like the numbers I expected. But I'm still surprised by your graph that seems to imply that when Vgk becomes more negative than -1.5V, that Ig rises to many tens of nA for the E180F. Or am I reading it wrong?

[TimFox]

I couldn't use log scales, since the current changes polarity as the grid voltage approaches zero.

Here's a summary of data taken with a range of tubes, triodes and pentodes.
I wanted to compare them at a normal operating cathode current value.

(Attachment Link)

Yes, it's a nuisance that grid current has a wide range and changes polarity. Your spreadsheet numbers are much more like the numbers I expected. But I'm still surprised by your graph that seems to imply that when Vgk becomes more negative than -1.5V, that Ig rises to many tens of nA for the E180F. Or am I reading it wrong?

I checked the original data, and the large grid current as the grid approaches cut-off did happen.  I’m not sure why, but I suspect ion current.  I’ll look into my other data.