Ultra Low Noise JFET

[G0HZU]

The image below shows a recent test setup I used to measure the noise figure of various BJTs at various bias points. It could also show 1/f noise. This setup worked really well and was very repeatable. I think the schematic below is correct. I still have the jig and the (E5052A signal source) analyser set up here and could adapt it for measuring JFETs. I have lots of JFETs here and I think I have some CPH3910 here. I would normally use these up at RF though...

I also have opamp measuring jigs and can measure the noise figure vs source resistance for lots of common opamp types including some really low noise types. I use the same analyser for this stuff and it has a really low noise floor

That is a nice simple setup, I expect that the CPH3910 would have a fairly high 1/f corner frequency of between 100Hz to 1kHz because of its RF nature, but any data you can provide in this regard would be appreciated. It is so valuable to have real world data from an independent source.

Thanks. The LT1677 acts as an impedance changer down to the 50R input impedance of the analyser. With the BJT removed and the 2700R collector resistor switched in, the LT1677 should only increase the noise level from the 2700R resistor a small amount. Many years ago I did a lot of receiver IF design using opamps and I have various spreadsheets that can predict the opamp noise figure vs frequency for a given resistive source impedance. There's lots of number crunching going on in that spreadsheet but it usually works really well across a wide range of opamp types.

I've added graphs to the spreadsheet to show the predicted noise level on the E5052A analyser and this is included in the image below. The spreadsheet uses the current and voltage noise (vs frequency) from the opamp datasheet and this can be used to predict the noise output level and noise figure etc and this is affected by the choice of source resistance and also the values of the feedback resistors around the opamp. The spreadsheet factors in the noise from these feedback resistors so the results it produces should be quite representative of the real opamp.

See the simulation plots below. There's also a measurement using the E5052A SSA and this shows the expected noise output from the opamp. The noise floor of the E5052A is really low. Much lower than the noise from the LT1677A opamp.

The 2700R resistor on its own should produce about 6.67nV/rtHz but the LT1677 degrades this to about 7.4nV/rtHz as in the simulation and the real measurement. As long as the BJT under test can produce significantly more noise than the test jig noise, the system noise figure measured by the analyser should be dominated by the BJT noise. It is possible to include secondary correction if necessary, but so far I've not needed to do this across a wide range of BJTs and source impedances and operating points.

[G0HZU]

In case it is of any interest, here's some old noise figure test results of various BJTs at 50 uA Ic and a 47k source impedance.

The dB numbers are for the noise figure of the BJT with a 47k source resistance and a collector current of just 50uA and the noise figure was measured at a few kHz.

BC337 1.2 dB (bandwidth limited to about 30 kHz at -3dB)
BC109C 0.52 dB
2N3904 1.15 dB
KSC1845 0.6 dB
BC107 0.72 dB
ZTX851 1.7 dB (bandwidth limited to about 3 kHz at -3dB)
BC550C 0.45 dB
BC849C 0.47 dB
2SC3324 0.68 dB
2SC2713 0.72 dB
BC547B 0.8 dB
2SC945 1.14 dB

There really isn't much in it although the ZTX851 wasn't really a contender here due to the reduced bandwidth it offers with a high source impedance.

At much lower source impedances, the ZTX851 performs the best in terms of noise figure. At higher collector currents some of the BJTs produce quite a bit of low frequency flicker noise. However, in the tests above at just 50uA, there was hardly any flicker noise visible, even with the BC107. The very same BC107 produced huge amounts of flicker noise at higher bias currents and a lower source impedance. Much worse than the other BJTs.

[G0HZU]

To make measurements like this requires some confidence in the test gear. So some time ago, I did do quite a few tests on the jig and the E5052A analyser to try and explore any limitations. The setup and test result below is just one of the automated tests I do on the E5052A to make sure it can measure signals without introducing significant errors due to FFT/sampling/windowing errors. It appears to use a peak detector when measuring cw signals and this explains the very flat frequency response result below, even with very small increments in frequency to try and weed out any hidden ripple errors. You can see the E5052A is very accurate in terms of flatness and absolute level when compared to the Keithley 2015 THD multimeter.

The 80 dB attenuator is a precision (and well proven) design, very accurate and very thermally stable and usable to several MHz.

[moffy]

That's quite a setup with a really low noise floor, no stray harmonics showing and flat response. That is lab grade quality, nicely done and thanks for sharing. Any further results you generate would be appreciated.

[G0HZU]

Thanks. I managed to find an old noise floor plot of (just) the E5052A with the baseband input shorted. I'm not sure how meaningful it is to short the input like this but it shows a very low noise floor when driven by an ultra low source impedance. If driven by a 50R source, the noise figure of the E5052A is typically under 3 dB. I think the B version (E5052B) is better in this respect and the noise floor down at 10 Hz is lower on the B version too.

But this is still good enough for the stuff I'm currently doing. I think this noise floor performance is much better than a lot of mainstream audio analysers and this allows the direct measurement of some really low noise devices that have a low source impedance. Eg opamps etc.

I'll try and get set up for testing JFETs but I may end up having to make a custom test jig for JFETs.

The E5052 analyser uses correlation to reduce its own internal noise and this is how it is able to achieve a really low noise floor. However, I'm not sure how accurate it is when pushing really deep like this. Usually, the testing I do means measuring noise levels above -160 dBV/Hz and the E5052A seems to be very accurate at noise levels like this.

[G0HZU]

The plot below is an old measurement of an LT1677 opamp set for unity gain direct into the E5052A. It shows -170 dBV/Hz which is just over 3.1nV/rtHz and this agrees with the datasheet. Generally speaking, the flatter and smoother the trace on the E5052A, the easier time it is having measuring the noise accurately. It can measure 3.1nV/rtHz really well, so any BJT (or JFET?) under test just has to provide enough gain to dominate the low noise floor of my test jig. Usually, there's about a 16-20 dB margin and this means very little in the way of errors due to the noise contribution from the test jig.

[moffy]

I just checked the price of some used E5052s and for only $30k-40k you can have a used one also! They list that they do 10MHz to 7GHz, but your trace shows baseband so I assume that you can bypass the demodulator and just connect the ADC stage directly to the input?

[G0HZU]

Yes, the E5052 also has a baseband input port on the front panel. On the E5052A this is a factory option that covers 1Hz to 40 MHz via a front panel BNC connector that sits just to the left of the main N type input connector. I think all E5052B support baseband across 1 Hz to 100 MHz. Very few E5052A models will have the baseband option enabled even though they have the baseband BNC connector on the front panel.

[mawyatt]

In case it is of any interest, here's some old noise figure test results of various BJTs at 50 uA Ic and a 47k source impedance.

The dB numbers are for the noise figure of the BJT with a 47k source resistance and a collector current of just 50uA and the noise figure was measured at a few kHz.

BC337 1.2 dB (bandwidth limited to about 30 kHz at -3dB)
BC109C 0.52 dB
2N3904 1.15 dB
KSC1845 0.6 dB
BC107 0.72 dB
ZTX851 1.7 dB (bandwidth limited to about 3 kHz at -3dB)
BC550C 0.45 dB
BC849C 0.47 dB
2SC3324 0.68 dB
2SC2713 0.72 dB
BC547B 0.8 dB
2SC945 1.14 dB

There really isn't much in it although the ZTX851 wasn't really a contender here due to the reduced bandwidth it offers with a high source impedance.

At much lower source impedances, the ZTX851 performs the best in terms of noise figure. At higher collector currents some of the BJTs produce quite a bit of low frequency flicker noise. However, in the tests above at just 50uA, there was hardly any flicker noise visible, even with the BC107. The very same BC107 produced huge amounts of flicker noise at higher bias currents and a lower source impedance. Much worse than the other BJTs.

Would be interesting to see how commercial discrete SiGe transistors perform wrt NF like the old BFR740L3RH for example. We know around ~2006 IBM 8HP/XP SiGe devices had exceptionally low NF (tho can't recall what Dr Sandy Weinreb at JPL/NASA measured with the 8HP SiGe devices we provided) as they replaced the LNAs in the Radio Telescopes in VLA in New Mexico. Believe the VLA LNAs that the SiGe devices replaced were based upon pHEMTs, which hints just how good the IBM SiGe devices were!!

Anyway, interesting NF data and nice setup

Best

[G0HZU]

I quickly powered up the test jig again and fitted a 2SC3324 low noise BJT to it and selected a 470R source impedance. I set the collector current to 1 mA and then took a noise plot with the E5052A. I then swapped the BJT to a NOS BC107 and biased it to the same 1 mA collector current. You can see the BC107 has lots of LF flicker noise compared to the 2SC3324. There is a trace of flicker noise with the 2SC3324 as well but it is only about a 2 dB increase which is similar to the datasheet when comparing noise at 10 kHz and 1 kHz.

Not all BC107s are this noisy. Some are quite good although they can't match the 2SC3324 for noise level above a few kHz.

The plot below shows the difference and this is why it is always important to consider (and measure) flicker noise. At higher collector currents this BC107 generates a LOT of flicker noise. It can get really noisy and it isn't faulty. It's just one of the noisier ones from the same batch in the same bag.

I'll try and measure some JFETs but it won't be for a day or so...

[G0HZU]

I can't really measure UHF+ rated BJTs in this test fixture because of the risk of instability. Fast BJTs can self oscillate up at UHF in a basic fixture like this.
When I have measured faster parts (eg Ft 5 GHz) there is higher flicker noise compared to the 2SC3324 for example. But I have to make sure the BJT doesn't oscillate in the fixture. It's usually really obvious when it oscillates but I also use a sniffer probe and another analyser to check for instability.  The BC107 isn't getting noisy due to instability, I've already ruled this out.

[moffy]

Can you specify the gain of the transistor stage in your setup so we can do an input referred noise calc? You have a couple of switches in your test jig that make that a little indeterminate.

[G0HZU]

I can fully set it up again with the Keithley 2015 THD and then measure noise figure automatically via GPIB using an excel sheet and this lists everything including signal and noise levels and gain etc. The system measures the input level accurately and also the output level and the output noise level.

In the meantime, I've crudely simulated the jig with a BC547B at 1mA Ic and a 470R source and this predicts an output noise level of 302nV/rtHz. This is about -130dBV/Hz so this agrees with the E5052A plot below. I've added a macro that converts across to nV/rtHz and the E5052A measured 307.6nV/rtHz which is close enough...

When I measure noise figure formally, the automated system measures the signal to noise ratio at the input and also at the output and computes noise figure from this data and this is a 1 click (automated) measurement. So the noise figure measurements should be fairly accurate. I'd like to make a neater and better test fixture and improve the automation yet further so this is probably going to be one of those projects that never ends...

[moffy]

Thanks for that, so the input stage gain would be roughly 100 (2.7k/26).
P.S. Most of that noise comes from the 470R @ 2.8nV/Hz^0.5, the 2SC3324 is pretty good.

[G0HZU]

If it helps, I just connected it all up for a 2SC3324 and pressed the run button (after entering the 470R source resistance and the 2700R collector load resistance).

The rest is all done automatically via GPIB including the collector current measurement. It measures it all and spits out an E5052A screenshot and does it all using Excel. The screenshot of the E5052A is taken a split second after the automated measurements so the marker data will be subtly different to the cells in the Excel sheet.

But you can see the noise figure of the 2SC3324 at 1mA Ic and 470R source resistance was measured at about 0.6 dB at a few kHz.

This seems to be valid looking at the datasheet. I've added a red dot for 1mA Ic and 470R source Z and the noise figure should be well under 1 dB. So it all looks good to me... However, I didn't measure the room temperature and I have just run it for the first time in ages. So I might have missed something in the setup. But there should be lots of data in the excel sheet to look over in terms of noise and signal levels

[G0HZU]

To speed the noise figure measurement up I start at 1 kHz and measure the noise dozens of times across a narrow range of frequencies near the 1.2 kHz test tone. I convert each one to a power then convert back to dBm/Hz to get the average noise power.

Doing it this way gives repeatable results for noise figure usually within +/- 0.1dB and it can do a noise figure measurement every few seconds like this. It still needs some work to make it easier to set up and the fixture needs to be remade as it looks a bit of a tired mess as this was made from joining together two older test fixtures.

So when I make a dedicated JFET fixture I should be able to measure the noise figure quickly. However, to measure 1/f noise below 1 kHz takes much longer, especially if I want to measure down to just 10 Hz. The E5052B is supposed to be much faster than the E5052A in this respect...

I've updated the fixture schematic so the test points align with the Excel spreadsheet and I've simplified the schematic to only show the 2700R resistor at the collector.

[G0HZU]

In my haste I've connected up the 34401A DMM leads backwards for the current measurement but this doesn't affect the result. But this is why the gm and current and Re show negative values in the excel spreadsheet. These numbers aren't used in the noise figure calculation.

[David Hess]

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.

[David Hess]

I can fully set it up again with the Keithley 2015 THD and then measure noise figure automatically via GPIB using an excel sheet and this lists everything including signal and noise levels and gain etc.

In the past I have made low frequency spot noise measurements and calculated the noise spectral density from them.

Given that the Keithley 2015 is has been discontinued for a while, what would you use in its place if you had to?

[G0HZU]

What is nice about the Keithley 2015 THD is that the BNC (sinewave) outputs are isolated/floating so this minimises ground loop (pickup) issues in the system. The only mains earth ground is at the E5052A baseband input. The DC output ports on the E5052A that provide the low noise Vcc and bias are also isolated and so are the PSU connections.

If the 2015 failed then I guess I could try using my Analog Discovery 2 to generate the test tone. I have other meters to measure the AF level at TP1. I really don't know how suitable the AD2 is for a task like this though. Otherwise I guess a modern Arb generator could be used if it has isolated outputs. I keep meaning to buy a decent 16 bit dual arb generator but I'll wait to see if any new models pop up soon. I'd want something decent so expect to have to dig fairly deep to get one.

[moffy]

@G0HZU, just a point of interest, but how is the 100uF capacitor and C4_big constructed, (might appear a bit dumb, but I'm interested how they are shielded if any)? Any photos of your test jig and setup would be appreciated, the connections etc are really important for such low level measurements.

[G0HZU]

C4 doesn't have a defined value on the schematic because this is really a select on test part. The value of C4 I choose depends on the source resistance value I select. R3 is similar as R3 and C4 affect how long the circuit takes to settle at the target collector current. For medium to high source impedances I tend to use a 470uF SMD tant (with low ESR at LF) for C4 and leave R3 at several k ohm. If I opt for a lower source impedance (eg 50 ohms or 100 ohms) I can plug in several more 470uF tant caps in parallel with C4 via sockets/pins. But this means the circuit takes longer to settle. I'm not really looking to measure Rbb, so the lowest source resistance I've used is 50 ohms. The 100uF cap is also a SMD tant part.

[MT4S301]

Will cap leakage current add to DUT low-freq noise ? Or it is generally small enough compared to Johnson noise of the 470R resistor?

[G0HZU]

What I found when developing this test fixture was that every component has to be considered as a source of noise and a lot depends on the choice of source impedance and the bias point in terms of how each component can spoil things.

I ended up using metal film resistors everywhere and some electrolytic and some tant caps. 1/f noise can creep in and dominate things and I'm still not certain my jig is OK. However, I think it is OK nearly all of the time above about 50 Hz when measuring BJTs across a range of source impedances and bias currents. For example, I can measure a ZTX851 and get a noise figure of about 1 dB or less at a few mA bias current and a 50R source impedance. Below 50 Hz there is a bit more 1/f noise than I expect to see but at the moment I'm not too worried about this.

I'm not sure about cap leakage effects and how this can generate 1/f noise that could dominate my test setup. I think my active filter on the Vcc supply is probably the weak link below 50 Hz.

With a low source resistance I have to make sure the reactance of C4 is really low below 50 Hz or this can cause issues. At the end of the day, it's knowing when to stop making improvements just for the sake of improvements and I feel I'm fairly close to that point now as I don't want to spend lots more time on this project. It isn't quite right below 50 Hz sometimes, but I'm not too bothered about this.

[G0HZU]

To show what I mean, I just tested a ZTX851 with a 50R source Z and about 4mA Ic.  The automated excel sheet showed a measured noise figure (using the E5052A) of about 0.5dB at a few kHz. Edit: I've updated the spreadsheet to include the correct temperature and this brought it down to 0.47 dB.

If I measure the noise down to 10 Hz with the E5052A you can see there is some flicker noise below 50 Hz. I think this is higher than it should be. See the screenshot below.

The other image shows a simulation using Genesys and the manufacturer's Gummel Poon model data for the ZTX851 under similar conditions. It shows a noise figure of just under 0.4 dB. This is quite close to my measured result and I haven't applied secondary correction for the added noise of the opamp and the E5052A. This may well add a tiny fraction of a dB to the measured noise figure. So I'm quite happy with results like this as this is giving good results even with a low source impedance