This investigation started as my attempt to find a replacement for one of the lowest noise JFET’s out there – BF862. It was a wonderful part designed for AM radio front-end. Process was really clean and well executed so this part had a noise density of 0.8nV/rtHz and a very low 1f noise corner with exceptional repeatability and resulting yield.
I had a stash of couple hundred BF862’s and thought that they will still be produced for some time to come. So I used them as a “placeholder” part basically everywhere, not worrying about availability. My rationale was abundance of Chinese AM radio stations. I thought – no way this part could EOL when there is still such a large market. Well, guess what – I was wrong. In a year or so of my sweet oblivion they went totally extinct.
I realized this only when I got a freelance proposal to design something very low noise and absolutely not audio related. I thought to my self – oh, this is gonna be walk in a park. I will just parallel bunch of my favorite JFET’s and job done. Better order some for sorting now, and BOM! There is none :/
My natural reaction was to order everything I could still find from the “usual suspects” as ebay and ali and soon I got my table full of BF682’s! Just one small problem. They were all fakes…
Wanna guess which one is which? Not a fun game, when you have about 1000 of these noise-generators on your table and deadlines are coming soon. So I posted in this thread as a last hope attempt and the search began.
I won’t expand here on all the semiconductor noise physics and low-noise design considerations. In order to understand and successfully replicate (if desired) my measurement setup, reader has to understand just the basic underlying principle: noise performance of a circuit is limited by the first amplifying stage. Sure, there is always ifs and buts in such a complex topic, but for our intents and purposes here – this has to suffice.
So essentially what we are doing here is building an amplifier using the device that we are measuring (DUT). Then feeding some known amount of input noise and calibrating our measurement setup. When we remove this input noise and measure again – we are only left with a noise generated by our DUT. Neat, isn’t it?
Couple words about above schematic. 3V battery and 200k pot is generating negative bias voltage for a desired DUT current. Battery type is not critical as the R1/C1 filter network will take care of any residual noise. R3 is our “100kOhms-thermal-noise-reference” which we will use to generate 40.5nV/rtHz noise base line. S1 shunts and effectively eliminates R1 when a measurement is made. Resistor type again is not critical as there will be virtually no current flowing through it. R2 together with 9V battery seems like a good compromise to bring most of the JFET’s to their linear region under 10mA of current. Practice (and theory) shows that running more then 10mA of current brings diminishing returns noise wise. S2 is disengaged when measuring this standing current and I would recommend removing any leads and multi-meter XMM1 from circuit when doing final measurement. They will pick up a lot of hum.
Amplifier A1 doesn’t have to be a particularly low-noise type. We only care about it having enough gain (about 60dB) and enough of bandwidth for our chosen measurement frequency span. I have used my own LNMA that I build in a half an hour or so many years ago (that’s why I’m reluctant to publish it :). Good example of such an amp would be Scott’s circuit from this thread.
Nothing is stopping you from using Scott’s circuit “as is” with it’s clever current-auto-set. I just didn’t had photovoltaic mos drivers at hand and couldn’t be bothered to order one.
Couple last notes. Setup must be re-calibrated every time DUT current is changed, because DUT trans-conductance (and circuit gain) will change accordingly. Keep DUT in it’s linear region. For BF862 this means at least 2-3V on the drain. Battery type for the DUT is important! Don’t use rechargeable battery’s here. 9V Alkaline seems to be the lowest noise per volt battery I have measured.
And last but not least – shielding. Nothing beats classics of the genre – cookie box! :] Except maybe vandalized tuner box inside a cookie box. Well, you get the idea. Generating clean nano-volt spectrum is not a trivial task at all.
There was a lot of independent noise measurements done and published over the years on BF862 noise spectrum. And it always comes to about 0.8nV/rtHz with varying degrees of 1f corner. So it’s safe to do a sanity check with this part.
Above you see measured spectrum right after calibration. Seems like pretty good match to declared 0.8nV/rtHz. Below are distribution of 4 random samples.
1f corner is about 100Hz. Sometimes you get noisier samples (like #2 above), but that’s a really rear one. What is important here is that they all bottom out at ~0.8nV/rtHz. Here is how noise spectral density scales down with current.
I’m NOT claiming that all my measurements below are accurate in absolute sense, but at least now I’m very confident they are accurate when referenced to BF862 noise. Which was my original intend.
First let’s see how current production JFET’s stack up in comparison to mighty BF862.
Seems like BF862 still holds ground as having lowest overall noise level, but 2SK2394 has much better 1f corner! Actually it’s best I have ever seen from a JFET or BJT for that matter, a really spectacular part! And it’s noise floor is maybe some 0.1nV/rtHz higher. I say maybe, because I must admit this falls into my calibration/measurement uncertainty. So for my intended applications 2SK2394 seems like a clear overall winner and my new favorite. I also have 2SK3557 on order. It looks like the same silicon as 2394 just lower trans-conductance, so I don’t expect any improvements. I will make an update when I measure them (don’t hold your breath though). Now let’s see how these modern parts look compared to older favorites like 2SK170.
NSVJ3910 seems like a good match for 2SK170. Almost identical noise spectrum’s, but it has twice as much trans-conductance (22 vs. 40 mS) and five times lower input capacitance! (6 vs. 30pF). So a huge overall improvement over now legendary 2SK170. Also, negligible NSVJ3910 input capacitance makes this part most attractive in terms of paralleling then any other JFET measured here. Just a word of caution – it was the ONLY part that decided to be ~350Mhz oscillator instead of amplifier in my measurements. So care must be taken to isolate gate appropriately for the application.
Most strangest result is the difference between NSVJ3910 and CPH3910. It’s suppose to be same silicon in different package, but CPH3910 has this hump from 100Hz to 9Khz. Now I’m not an semiconductor process expert, but it looks like some manufacturing impurity to me. This could be just my batch (Farnell 2020.06 stock) or it could be how this part really is. I made some distribution measurements to see if this is consistent defect in my batch.
Above is a noise spectrum of 10 random samples from a batch of 40 pcs. Seems like 70% of samples are affected to some degree. For comparison: below is 10 samples of 2SK2394 (from a batch of 40).
Almost ideal parameter spread except one “fluke”. This just shows again that for a really noise-critical applications, hand selection is the only way to go.
These measurements were done when I was testing and calibrating my measurement setup for the first time. I was getting some erroneous results with very high 1f component so I decided to measure the battery I was using. They were done with just my average LNMA which has a noise floor of some ~1.5nV/rtHz (black line in a graph below).
I also measured some Li-Po 8.6V packs and some Eneloop sticks, unfortunately didn’t saved results. But the trend is clear – for lowest noise voltage source rechargeable battery’s seems like a bad choice. My guess this is because of different chemistry. In theory battery noise should be equivalent to it’s internal resistance thermal noise. Which would suggest that a beefy 18650 capable of supplying 20Amps should have clear advantage over wimpy 9V Alkaline. But in practice, when current flows from 18650 there is a lot of 1f noise generated. It looks something like pop-corn noise in transistors in a sense that a large amount of charge carriers are accumulated somewhere in internal structure and they all decide to jump together randomly. This is clearly visible with faster FFT times and no averaging. Of course this might be battery manufacturer, charge level and wear dependent (used Panasonic battery’s with 50% charge were depicted here). Also chemistry could play decisive role as this paper suggests Ni-Cd having ultra-low noise. So overall this seems like a good subject for some thesis, unfortunately no further investigation was carried out as I’m not particularly interested in publishing papers :]
Very often in audio and in general electronics there is a need for low noise voltage reference. Series connected LED’s are often used as cheap, readily available and reasonably low noise alternative to a dedicated parts. But how much noise do they really contribute?
Here is a most typical example of plain (read cheap) diffused green LED. You can buy 100pcs. of them for 1USD shipping included these days. As can be seen from above spectrum, noise scales down with current. So running them hot but still safe at 10mA or so will get you down to 7nV/rtHz in lower region. This can be attenuated even further by simple RC filter.
Going for bigger package (5mm) worsen things slightly. Again, I suspect this will be manufacturer/model depended. There are 5mm LED’s that only belongs in a recycle bin.
Above is a good example of what I’m talking about. These old 5mm LED’s almost don’t produce any light output when compared to modern examples and are noisy as hell.
Red LED’s doesn’t bring any benefits noise wise and they also have lower forward voltage. Again it looks like one should avoid bigger 5mm packages. Let’s look what modern SMD packages bring to the table.
There is no way to give a meaningful general advice on SMD parts as these are very package and manufacturer depended. But there are really good parts like KPT_3216 from Kingbright. These green LED’s at 10mA measures just above the noise floor at about 4nV/rtHz at most extreme 1f region.
Not often diodes are used in applications where noise is of any concern. But when they are! …. no one knows what noise contribution to expect as this parameter is defined only for RF parts. They also can be used as voltage references so all measurements below were made using 4 diodes in series to give a better estimation when compered with LED’s.
Last graph suggest that four 1N4148 could give slightly more voltage drop at 10mA current and provide less noise then 3mm Green LED. Of course for larger voltages number of diodes becomes quite unattractive when space is considered.
For larger reference voltages, zener diodes become lucrative alternative instead of LED’s or diodes. There are a lot of horror stories (especially in audio) about zener diode noise. Let’s see if they hold any ground.
And it seems those stories are true. They don’t have any 1f noise, but shot noise is substantial and very current dependent. It’s just a way these parts work – they are reverse biased junctions operated in a break-down region. Takeaway here would be this: don’t use zeners in noise sensitive applications. If you really must use them – series connect lower voltage ones (up to 5V1 seems to be fine) and run as much current as package can dissipate.