My preamp used 6 InterFET IF3602 dual JFETs. 4 duals were paralleled as the amplifier and two duals were paralleled as a cascode load. Out of 25 JFETs I found ten that were suitable for my preamp. All the JFETs were characterized to and selected to match their pinch off voltages as closely as possible. Then the lowest noise JFETs were selected for the amplifier, the rest being good enough as loads.
Last edited:
There are several reasons why these JFETs were chosen.
I think you would be hard pressed to find any JFETs with noise as low as 350nV/rt-Hz at 30Hz. I've tried several other "low noise" FETs and none of them even came close to the performance that I needed. Then there's the operating current requirements.
The high gain of the preamp also necessitated operating the JFETs at a drain current of about 15mA per JFET in order to raise the transconductance (gm) high enough to get the required gain. These JFETs have a minimum IDSS of 50mA. The active load for the preamp had a resistance of about 100 ohms. That means the gm had to be about 1.6mS to for a gain of 44dB.
The impedance of the active load was low for two reasons 1) to keep the load from adding too much noise (<1.4nV/rt-Hz) to the total output, and 2) to keep the effect of parasitic capacitances from reducing the bandwidth to less than 10MHz.
Even if you could build the preamp for less, unless you have the right test equipment it's hard to see how you would verify the performance. I wound up building two of these preamps and only one had the noise floor of 300pV/rt-Hz, the other was closer to 500pV/rt-Hz, still pretty low by any measure. The matching in the high noise preamp was not as tight and may have contributed to the excess noise.
Last edited:
Paralleled BF862s. About 22 cents each in 100 quantity. Cascode them with paralleled higher-pinchoff devices to reduce the 862s' drain voltages and eliminate most of the Miller effect. If gain accuracy and stability is critical, use series feedback to a small resistor in the composite sources of the 862s, say one ohm (129pV/sq rt Hz at 300K).
There will be the usual bandwidth versus gain tradeoff, but the open-loop GBW can be very high. I understand Wurcer has a great article coming in Linear Audio about JFETs although I haven't seen the details. Danyuk has done some good work characterizing the 862 compared to other JFETs. He didn't look at the Interfet parts, but compared to many others the 862 shines.
Yes, not too many easy ways to characterize up to 10MHz. Not usually a big concern in this forum. I can understand why the ADI regulators required it (they even say RF in the description).
Brad
When I was looking for JFETs I looked at the BF862 saw that it only noise data at 100kHz. To most designers, this usually means that a parameter with limited or no data/curves has not been characterized nor is monitored during manufacturing. I have been bitten in the past by using a transistor that had excellent noise performance only to find out later that the fab had been moved and that the transistor no longer behaved the same because the noise was not a controlled parameter. This preamp had to be reproduced and maintained for the next 10 years at another manufacturing site and I could not take risk of designing in a transistor with uncontrolled noise characteristics.
The RF guys kept pushing me for noise data at higher frequencies but 12MHz was about as far as I could get. In order to make the noise calculations simple, the gain had to be flat (+/-0.25dB) from 10Hz to 10MHz. to keep the error to less that 5% over the entire bandwidth. I didn't want to have a bunch of gain correction factors since I was gathering about 500 data points. I agree that 10MHz is WAY beyond what most folks in this forum care about, they would probably be happy with 100kHz bandwidth where it's easy to achieve gain flatness of 0.1dB or better.
I should also mention that the preamp did not use feedback mostly to avoid issues with noise in the feedback path.
glenn
Last edited:
Hi,
I got 220 pV/sqrt Hz from averaging 20 ADA4898 op amps. It is closed loop, accuracy
and temperature independence come from feedback and it simply behaves as calculated.
Noise performance can easily be verified by measuring the thermal noise of resistors.
The drawback is that the bias system is medium impedance and one needs a large
input coupling capacitor if one wants a lower corner of 0.1 Hz. After intensive fight with
myself I have ordered some wet slug tantalums yesterday.
(AVX seems to be cheaper than Vishay, but still OMG.)
I'm also experimenting with a flock of IF3602s, that won't be cheaper. My signal source
can live with a Cin of a nF or two, but 1/f is better with the 4898. The only real
advantage would be the polarity independence and the harmless current surges
through the small coupling capacitor. Vcc must be ultra-clean, even
a LT3042 needs some help. Maybe I can buffer the 2nd stage with a BUF634 or so
and feed the pull up from there. Should give some PSSR.
With the ADA4898, PSSR comes for free.
regards, Gerhard
I got 220 pV/sqrt Hz from averaging 20 ADA4898 op amps. It is closed loop, accuracy
and temperature independence come from feedback and it simply behaves as calculated.
Noise performance can easily be verified by measuring the thermal noise of resistors.
The drawback is that the bias system is medium impedance and one needs a large
input coupling capacitor if one wants a lower corner of 0.1 Hz. After intensive fight with
myself I have ordered some wet slug tantalums yesterday.
(AVX seems to be cheaper than Vishay, but still OMG.)
I'm also experimenting with a flock of IF3602s, that won't be cheaper. My signal source
can live with a Cin of a nF or two, but 1/f is better with the 4898. The only real
advantage would be the polarity independence and the harmless current surges
through the small coupling capacitor. Vcc must be ultra-clean, even
a LT3042 needs some help. Maybe I can buffer the 2nd stage with a BUF634 or so
and feed the pull up from there. Should give some PSSR.
With the ADA4898, PSSR comes for free.
regards, Gerhard
Last edited:
Danyuk's JFET measurements
Measurements Rate SMT Low-Voltage n-JFETs Under Consistent Conditions | Power content from Electronic Design
Note that to avoid any self-heating of consequence the DUTs were run at 1mA and 2.5V Vds. We know that JFET channel noise has a weak dependence on drain current, about a reciprocal 1/4 power, so at an 862 typical Idss around 15mA we might expect nearly a factor of 2 improvement.
The 1/f corner is fairly low, although beaten by good bipolars.
EDIT: of course these data do not guarantee that other parts are comparable. If I included them in a design that anticipated significant volumes, I'd buy a whole bunch. Wayne of Pass Labs said he has a knack for picking parts that are almost immediately discontinued, so they buy production quantities from the outset
Measurements Rate SMT Low-Voltage n-JFETs Under Consistent Conditions | Power content from Electronic Design
Note that to avoid any self-heating of consequence the DUTs were run at 1mA and 2.5V Vds. We know that JFET channel noise has a weak dependence on drain current, about a reciprocal 1/4 power, so at an 862 typical Idss around 15mA we might expect nearly a factor of 2 improvement.
The 1/f corner is fairly low, although beaten by good bipolars.
EDIT: of course these data do not guarantee that other parts are comparable. If I included them in a design that anticipated significant volumes, I'd buy a whole bunch. Wayne of Pass Labs said he has a knack for picking parts that are almost immediately discontinued, so they buy production quantities from the outset
Last edited:
Hi,
I got 220 pV/sqrt Hz from averaging 20 ADA4898 op amps. It is closed loop, accuracy
and temperature independence come from feedback and it simply behaves as calculated.
Noise performance can easily be verified by measuring the thermal noise of resistors.
The drawback is that the bias system is medium impedance and one needs a large
input coupling capacitor if one wants a lower corner of 0.1 Hz. After intensive fight with
myself I have ordered some wet slug tantalums yesterday.
(AVX seems to be cheaper than Vishay, but still OMG.)
I'm also experimenting with a flock of IF3602s, that won't be cheaper. My signal source
can live with a Cin of a nF or two, but 1/f is better with the 4898. The only real
advantage would be the polarity independence and the harmless current surges
through the small coupling capacitor. Vcc must be ultra-clean, even
a LT3042 needs some help. Maybe I can buffer the 2nd stage with a BUF634 or so
and feed the pull up from there. Should give some PSSR.
With the ADA4898, PSSR comes for free.
regards, Gerhard
I get that you can use the thermal noise of a resistor to "calibrate" the RMS noise of the amplifier but you can't determine the noise spectral density with that method. For my work, the noise spectral density as a function of frequency was just as important as the RMS noise. I was able to show that from 10Hz to 10MHz, the noise spectral density of the preamp was <20% of the noise of the LDOs under test and thus the error due to the preamp was <4%
The RF guys don't care about RMS noise in their calculations for phase noise. Phase noise is determined for a specific offset frequency from the carrier frequency. In this way they can determine how much phase noise will result for a certain amount of noise at a specific frequency.
The same is true for the high speed (>2GHz) ADC/DAC guys. They want to be able to calculate the ENOB (effective number of bits) with a given amount of noise within the bandwidth of interest. These high speed ADC/DACs have virtually no PSRR and any noise on the supply rails gets translated into the baseband noise.
I am well acquainted with the ADA4898 (one of my favorite opamps) being a former ADI applications engineer. Unfortunately for me, the ADA4898 has a bandwidth of only 10MHz so I couldn't use them in the gain stages, remember, my preamp had to have at least 10MHz bandwidth. If I limited the bandwidth to only 100kHz and increased the gain, I was able to decrease the input referred noise to less than 150nV/rt-Hz. This wouldn't really be useful since the RF guys wanted noise data above 100kHz. ADI was the first LDO manufacturer to provide noise data from below 10Hz and above 100kHz.
I order to be able to use a reasonably size input coupling capacitor (10x 1uF/50V polycarbonate), I used a 10Mohm gate resistor on the JFET. Since I was testing LDOs, the low output impedance of the LDOs made the 10Mohm resistor noise irrelevant. Have you considered polycarbonate motor starting film caps for your coupling cap? I've used them on another project with good results.
You really have to watch out for excess current noise in tantalums especially wet slug types. I had one preamp that had a noise level that was 10x higher than one built with exactly the same components. I even swapped out the JFETs and the noise remained high. It turns out one of the solid tantalum capacitors had excess noise presumably due to leakage current. Once I swapped out the bad cap, the noise returned to normal.
It's true that the Vcc for the JFET gain stage must be extremely clean. I used the ADP7102 to set the voltage and two stages of capacitance multipliers to knock the noise down to about 10uVrms over the 10MHz bandwidth.
That's probably true given their low noise at 100kHz but you can't guarantee that the fab location won't change. It's been my experience that when fabs change, many times those unspecified parameters change as well. I've even seen issues when the fab location did not change but the wafer fab changed from a 4" wafer line to 6" wafer line and things went wonky.
In this case though, the preamp would only need spares and buying a bunch of these JFETs might make sense. However, if someone had to resurrect the design from first principles and did not have access to the tribal knowledge that these particular JFETs had excellent noise performance they might take a totally different design approach...
Taking advantage of unspecified parameters is risky at best and really only suitable for one off and hobbyist applications. You don't ever want to be in the position of defending your design decisions when they rely on unspecified parameters.
Measurements Rate SMT Low-Voltage n-JFETs Under Consistent Conditions | Power content from Electronic Design
Note that to avoid any self-heating of consequence the DUTs were run at 1mA and 2.5V Vds. We know that JFET channel noise has a weak dependence on drain current, about a reciprocal 1/4 power, so at an 862 typical Idss around 15mA we might expect nearly a factor of 2 improvement.
The 1/f corner is fairly low, although beaten by good bipolars.
EDIT: of course these data do not guarantee that other parts are comparable. If I included them in a design that anticipated significant volumes, I'd buy a whole bunch. Wayne of Pass Labs said he has a knack for picking parts that are almost immediately discontinued, so they buy production quantities from the outset
I remember reading this article and thinking, WOW someone else is in interested in this stuff. I also thought, Gee, I wish the author had taken data out to 1MHz or higher. Interesting article though, for as far as it went...
I eliminated bipolars since the relatively low input impedance would have precluded the use of my relatively small input coupling capacitor. ADI's MAT14 are especially good candidates.
What Harman would do to make sure, in volume apps for automotive, was to get the manufacturer to guarantee critical specs and give them a specific part number. This satisfied the QA people and worked well most of the time.
Agree, that would be most wise. The automotive industry is a good example where significant volumes exist for a long period of time. Unfortunately getting Chinese manufacturers to guarantee anything is tricky at best. That is unless of course you are talking about tens or hundreds of thousands of units per year, even then the price may not be attractive.
The reason why I don't like to rely on unspecified parameters is explained below:
Let me preface this by saying that I am currently designing rad-hard DC-DC converters for satellite applications.
Several years ago, someone (NOT me) developed a design that relied on a particular opamp that was screened for radiation hardness, initially with good yields. Things were good for many years. Then the opamp fab moved and now only 1 out of 4 lots pass and there is no guarantee that the yields won't drop. The testing costs about $20K for each lot so basically we are spending $80K, not including our engineering time and paperwork, to get one lot to pass. Each lot has about 1000 parts so the testing adds about $80 to each opamp. The poor yields have caused production line shut downs on several occasions and resulted in angry customers who missed launch windows.
There are opamps that are guaranteed to be rad-hard for about $200. In the long run it would have been better to procure a guaranteed part and factored in the higher cost from the start.
Unfortunately the original engineer is long gone so we can't flog him for getting us into this pickle ;-)
Last edited:
Of course, I can get noise density calibration from thermal noise, and that
means from 1st. principles. There are few things that deliver 1nV/sqrtHz at any
frequency like a 60 Ohm thin film SMD resistor. And the beauty is, you don't
have to mess around with carriers and FFT bin widths.
I have written a program that controls an Agilent 89441A vector signal analyzer
over the LAN to do FFTs over 7 decades, collect the data and combine them
to a single plot from 0.1 Hz to 1 MHz. It takes about 6.5 minutes for that range
at 100* averaging, with some overlap.
The lion's share of that time goes to the lowest decade.
The writeup of the preamp is at
< http://www.hoffmann-hochfrequenz.de/downloads/lono.pdf >
I still use the version with the 20 * 10uF WIMA foil capacitors
most of the time, but that has limitations below a few Hz. (the low
source impedance must short the noise of the bias-R. and that
needs a fat capacitor at 0.1 Hz.)
It is remarkable that the ceramic capacitors did not produce any problems.
Oh, I AM a RF guy. That effort would be wasted for audio, even if THEY
want to crucify me. And I would use those fast ADCs if they were space qualified.
It would make our system design easier.
Last edited:
Ah, you neglected to mention you used a vector signal analyzer. I assumed (my bad) that you were making RMS measurements for the entire bandwidth. I found that the frequency bin width for the E5505A resulted in fairly long integration times especially at the lower frequencies as you noted. I short the input of my preamp and measure the resultant input referred noise. Since I am measuring the output of an LDO, a shorted input simulates the impedance seen by the preamp at DC.
The E5505A still has the lowest noise floor of any instrument still in production even though the basic design is over 20 years old. The software driving it is pretty primitive by today's standards. It sort of runs under Windows 7 and has a tendency to lock up occasionally. I basically exported all of the data and dumped it into Excel to do the decimation and averaging.
The E5505A is apparently is very popular with the signal intelligence community. The waiting list for one of these is about 6 to 8 months. Keysight can barely keep up with the demand.
Last edited:
Yes, we are using an E5502B for oscillator phase noise with the dual reference
and cross correlation option, and I also had limited access to a R&S FSUP.
Given the free choice I would probably take the FSUP, although it is slightly more
pessimistic.
The 89441A is my own toy and it is more oriented towards recognizing all
cell phone modulations. It takes a lot of ext. preamp gain to
overcome its own noise, especially 1/f. But it can do Cross FFTs, gain &
phase from DC up etc and it is cheaply available because the telecom companies
need new boxes for LTE. I was told that mine was pre-owned by Motorola.
regards, Gerhard
and cross correlation option, and I also had limited access to a R&S FSUP.
Given the free choice I would probably take the FSUP, although it is slightly more
pessimistic.
The 89441A is my own toy and it is more oriented towards recognizing all
cell phone modulations. It takes a lot of ext. preamp gain to
overcome its own noise, especially 1/f. But it can do Cross FFTs, gain &
phase from DC up etc and it is cheaply available because the telecom companies
need new boxes for LTE. I was told that mine was pre-owned by Motorola.
regards, Gerhard
That summary gives me some idea of what I don't know !
Thanks for the insight.
I'd love to find that chip in an SMD package. Right now the little foil leads are a production problem.
