Just for fun: a superreg with <12 discretes??! ?! ?

Status
This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.
During a discussion about an add-on for upgrading 317/337 regulators, a question has surfaced:

Would it possible to design a "superreg" using only 12 components or less, preferably all discretes?

Here is an answer (to be taken with of pinch of salt, and in the spirit of Christmas magic, something that participants to the previous thread seem to have missed).

This example uses 11 discretes, and outperforms a cap-fitted 317 on all key parameters (without using any cap).
Does it qualify as a superreg?

Certainly not by today's standards, and the DC stability is approximate because of a crude, first-order temperature compensation, but the central question was noise-related, and it does not perform too poorly in this regard (and there is a lot of room for improvements, since the circuit is based on the <12 components constraint, and no cap).

May I remind everyone that this site is named DIYaudio, and members like to fiddle with creative, cheerful and unorthodox solutions (which is not synonymous with worthless), which is why I post this circuit.
Grinches, etc., please don't look further and move along....

Here is the PSRR of the basic circuit: ~80dB

724106d1545600050-fun-superreg-12-discretes-suprpsrr-png



This is the output impedance: ~1.6mΩ

724107d1545600050-fun-superreg-12-discretes-suprzo-png


The measured noise, in a 10Hz to 10kHz bandwidth varies between ~2µV and 4µV, depending on the LED used and the error amplifier transistor.
The most favorable combination (I only made a limited number of tests) is an old Siemens green LED (CQ-something) and a BC337-40 resulting in ~2µV, and if the LED and transistor are from 6N139 coupler (an attractive solution because of the thermal tracking, the single component and the good temperature compensation), this becomes ~2.5µV.

It is possible to compensate some of the parasitic parameters like Early effect to the first order by adding some tweaks.
The practical range of this type of tweak cannot exceed one order of magnitude, ~20dB: beyond, the adjustment become too sensitive to be of practical use, but for DIY applications, 20dB is realistic.
Here is the result with tweaks active (becomes >12 components, of course)/

724108d1545600050-fun-superreg-12-discretes-suprpsrt-png


724109d1545600050-fun-superreg-12-discretes-suprzot-png


Note that R3 is not a physical component: it simply materializes the order of the components on the output track: R5 needs to be upstream of D1.


It is important to note that the circuit is not self-starting: this can be seen as an advantage or a drawback, depending on the application.
In a previous example, I have used this as a "feature".

Disclaimer:
The circuit in its present condition is suboptimal in many respects: its main purpose is to deliver nice figures within the 11 discretes constraint, and although it works (I had to make actual measurements on the breadboarded version), I certainly do not recommend using it in its raw state.

It could become really useful with only cheap and minor amendments/additions, but it will probably remain unable to beat the best silicon $£€ can buy, made by the top foundries.

I will discuss later the practical adjustments required if the circuit has to be used in the real world, by people that do not necessarily have access to Farnell, Mouser, etc. and for whom a 317 is already a kind of luxury (2/3rd of the world population?)
 

Attachments

  • SupRpsrr.png
    SupRpsrr.png
    91.4 KB · Views: 3,375
  • SupRZo.png
    SupRZo.png
    91.4 KB · Views: 2,404
  • SupRpsrT.png
    SupRpsrT.png
    93.6 KB · Views: 2,395
  • SupRZoT.png
    SupRZoT.png
    89.3 KB · Views: 2,377
Last edited:
I posted the output impedance graph in an unpractical format.

Since editing of the attachments doesn't seem to be allowed, here it is in a more legible form (if a moderator feels like it, he/she can replace the file and delete this post).
Also included is a pic of the test setup

A clarification:
The picture shows the optocoupler transistor in the cascode role: I made this test, to check whether it would improve the noise, but it didn't, so the error amplifier location is the best possible, both for temperature compensation and noise (irrelevant).
 

Attachments

  • SupRZo.png
    SupRZo.png
    103.9 KB · Views: 376
  • SuperrPic1.jpg
    SuperrPic1.jpg
    444.3 KB · Views: 429
Last edited:
I have tested the "lowest noise version", the one using two large caps at critical locations (C2 and C3).

Now, the noise originating from the LED is mostly eliminated, as is the noise caused by the resistance of the feedback divider, and the noise gain is brought to 0dB.

The only significant cause of remaining noise is now the transistor

Unsurprisingly, all of this greatly improves the noise performance: it is now lower than 0.4µV for the 10Hz to 10kHz bandwidth.
I am presently unable to tell how much lower it is: I have reached my measurement limit.

I may be able to give more information in one or two weeks.

I tested both the BC337 and the 6N139 transistor as error amplifier.
As can be expected, the BC337 is marginally better, but so marginally that an additional discrete doesn't seem worth the trouble, based on the present data.

This is an interim result; it may change when I refine my setup (and ultimately, the noise performance will inevitably be dictated by this transistor)
 

Attachments

  • Supercap.png
    Supercap.png
    57.5 KB · Views: 1,674
bc337 replacement

the dominant noise will be voltage noise,of both emitter and base resistance.

as standard NPN's have a thin base for high gain, and the base is made of P-material, base resistance is usually higher for NPN's. RF transistors are optimized in geometry for low base resistance.
I found the BFU530A with 0.9dB NF, a good possible replacement candidate..
 
We have been using a version of this for our headphone amps for the last year :
Voltage Regulator From Discrete Components

Of course not comparable to the Jung/Didden in pure specs.
But it is really simple (9 components without current limit), good stability, and very low noise.
Noise is basically 1x Vbe of 2SC3324/2SA1312 + 1x HLMP6000.
The latter is known for its very low noise.
(Photo is from an older version using LM329 instead of the LED.)
The current source is key to PSRR performance.

Perhaps someone can come up with an even simpler one ?


Happy New Year,
Patrick
.
 

Attachments

  • SDR LED PSRR Freq Public.asc
    6.6 KB · Views: 212
  • 171001 SDR Frt.JPG
    171001 SDR Frt.JPG
    84.6 KB · Views: 673
Those simple discrete regulators come basically in two flavors: one is the follower output, as shown by Patrick, the other is the common-emitter output.

Both have their peculiarities and advantages: the common-collector is self-starting and more familiar, but the C.E. is LDO by nature, and it has the potential to provide better performances, since the reference voltage will generally be created from the output.
This however makes it non-self-starting. This can be seen as a liability but also as an advantage.
I generally manage to make an advantage of it, like in this example:
A PSU controller on a shoestring

Other starting methods are possible: completely self-starting, with a pre-bias or a capacitive kick-start for example, with two push-buttons, under the control of a µcontroller, etc

Another advantage of the CE topology when used with a LED as a reference is the first order temperature compensation of the Vbe by the LED.
With a 6N139 coupler, the compensation is very close to ideal.

It is in general preferable to actively bias the reference element (LED) rather than just relying on the "natural" current.
In the CE topology, this is inherent, but it can be added in the CC configuration: in Patrick's circuit, this result in ~5dB improvement for PSRR and internal impedance.
It costs one more resistor of course....
 

Attachments

  • PatSuReg.png
    PatSuReg.png
    125.1 KB · Views: 586
Other starting methods are possible: completely self-starting, with a pre-bias or a capacitive kick-start for example, with two push-buttons, under the control of a µcontroller, etc
Have you consider making the start-up by generating a start-up pulse?
It only takes
- a capacitor, like 1uF
- a resistor to set the time, like 100kOhm
- a diode, like 1N4148 to block further current when discharged

I have used this method frequently when using common-emitter output regulators
and when supplying reference from the output.
I have seen others using this.
 
Here is the first real-life application of these regulators:
Sanity-check + end-result: a x 1000 measurement preamplifier

They are used inside the low-noise wall-wart supply.


The main adaptations are the inclusion of a permanent startup circuit (R11, R12, D5) and the splitting of the cascode resistor into R5 and R10, to give Q2 more room at low dropout voltages.
I didn't include LED bypass capacitors or Early tweaks, as the performance was already more than sufficient
 

Attachments

  • supRegLNA.png
    supRegLNA.png
    72.4 KB · Views: 1,939
724451d1545770831-fun-superreg-12-discretes-supercap-png


Unsurprisingly, all of this greatly improves the noise performance: it is now lower than 0.4µV for the 10Hz to 10kHz bandwidth.
I am presently unable to tell how much lower it is: I have reached my measurement limit.

I may be able to give more information in one or two weeks.

I have been able to refine the measurement, now that I have a proper measuring gear: the noise in the 10Hz to 10kHz range is revised to 0.25µV, but this does not represent the ultimate limit: the circuit is still in the breadboard form, and most of the visible noise are bursts from the PLC, pervading everything.
The second contributor is 50Hz/100Hz + harmonics induced from the environment.
"True" noise is thus somewhat lower, probably in the 0.15µV region
 
The schematic in post #11 confuses me a little bit.

  • Is the base of transistor Q5 actually connected to the regulator output node?

Yes, it is, for the reason given by Basreflex.
However, it leaves little breathing space for the transistors, and it demands a low Vto from the PMOS to be able to operate in ~LDO mode, which is why I opted for a divider in the practical application I used it for:

729521d1547660752-fun-superreg-12-discretes-supreglna-png


  • [*]Does any current flow in resistor R3? Its two terminals appear to be shorted.
It is one of the open-loop enhacement tweaks: when the short is removed, this small resistor (it could be a few mm of track) creates a negative resistance that compensates the residual output impedance.

It can be made larger, to pre-compensate the downstream wiring resistance for example.

R1/C1 forms the open-loop ripple compensator when connected to the emitter of Q5.

All that is explained in the first post
 
Yes, C2 is actually 47µF, and has a 100V WV, which might look vastly excessive, but the lowish capacitance resonates (slightly) with this particular transformer's leakage inductance and mandates a high enough voltage rating to cope with the reactive role, something borderline for standard Ecaps.

This value just balances the the two supply polarities with the drain imposed by the LNA.

These values only apply to this particular transformer, used in these particular conditions, but once the principle is understood, it can be applied elsewhere.

While we are it, the role of D6 (which apparently does nothing useful) is to prevent a polarity inversion of C2 during random, unfavorable cold-start conditions.

D1 and C4 are connected correctly: C4 (very close physically to the diode) eliminates possible switching noise due to snap-recovery (relatively rare nowadays in modern slow diodes), and more importantly prevents it from acting as a PIN diode, imposing a 50Hz modulation to the RF haze pervading our modern environments.
This supply was custom-built built for a sensitive LNA, which explains those precautions (and no, I am certainly not a paranoid crackpot chasing non-existent problems, I am 100% objectivist, and these precautions are a result of real-world, measurable experiences).

This pic illustrates the operation of rectifier part of the supply, and I also include the asc if you want to check by yourself


attachment.php
 

Attachments

  • LNAsupp.png
    LNAsupp.png
    73.6 KB · Views: 1,116
  • LNAsupp.asc
    2.3 KB · Views: 55
I have built another test prototype, and begun some tests.

This is the test circuit:

attachment.php


First, I am going to make the tests without any capacitors, just like a monolithic regulator.
I had to add a small 68pF compensation capacitor, as the circuit was marginally stable: sometimes, oscillations appeared depending on the input voltage, load, etc.
It is a plain-vanilla version, using completely ordinary components: unsorted BC337 and BC548 (no suffix), an ordinary green LED straight from my drawers, etc.
Input and output bypass caps are present, except for the PSRR measurement.

I built it in realistic form-factor, mimicking a regular VR: GND connection in the middle, input and output on the sides.

attachment.php



It is configured for 12V; the actual output with the values shown is 12.5V.
At 500mA, it requires 1.8V I/OΔ to regulate, very much like a conventional regulator.
It has the capability do better: the dropout is caused by the generic BS250 from Multicomp.
With a better PMOS, a dropout of <1V would be possible (using a blue LED as a reference would have the same result).
Thanks to its discrete nature, it also has a somewhat larger current capacity: I tested it at up to 4A (that's the limit of my lab supply).
The board has a provision for an output impedance compensation resistance (jumper receptacle), but I had to short it: its residual resistance resulted in a -60mΩ output impedance!

I made the tests at 500mA, I/OΔ=5V.

The total output noise (10Hz-10kHz) is 6.5µV.
This is relatively large, but it includes the noise of the transistor, and of the unbypassed LED and resistive divider, multiplied by the regulator gain (~10).
"Cheat" capacitors should improve the situation

The output impedance is 6mΩ @100Hz, 6.3mΩ @1kHz, 8mΩ @10kHz and 60mΩ @100kHz.
The figures are better than the sim, probably because the large blob of solder I used to short the compensating resistor has some residual resistance.
When the compensation is used, the output impedance can be cancelled completely (or even made negative).

The PSRR is -80dB @100Hz, -63.1dB @1kHz and -42dB @10kHz.

Next step will be to add the capacitors.
 

Attachments

  • 12DSr1.jpg
    12DSr1.jpg
    366.5 KB · Views: 881
  • 12DSr0.png
    12DSr0.png
    19.2 KB · Views: 909
I have added the bypass capacitors:

attachment.php


attachment.php



Now, the 10Hz to 10kHz noise is <0.25µV. I say "<", because part of the noise is visibly caused by PLC pulses intruding, and hum induced in the measurement cabling.
The output impedance is 300µΩ @100Hz, 320µΩ @1kHz, 340µΩ @10kHz and 850µΩ @100kHz.

The PSRR is -109dB @100Hz, -106dB @1kHz and -87dB @10kHz.

The compensation cap is not required anymore with the bypass caps, but it doesn't seem to harm the performances either.

For the moment, the circuit is not self-starting: I need to start it up manually.
Various types of starting circuits have been discussed previously; which one to use depends on the exact application
 

Attachments

  • 12DSr2.png
    12DSr2.png
    20.8 KB · Views: 851
  • 12DSr3.jpg
    12DSr3.jpg
    330.5 KB · Views: 850
I just realized that I made a mistake: I had moved the test setup and the essential test instruments to the quietest (electromagnetically speaking) place of my house, to get rid of the various interferences and make a "clean" measurement.

It did improve matters, but I noticed a large VLF contents in the noise, like huge jumps.

It then dawned on me that the LED bypass cap (560µ/4V) was a solid polymer type.
I used it because it was compact, and perfectly suited in size and voltage.

Unfortunately, these caps have a significant leakage current, and this current varies randomly.

I removed it, and the noise dropped to <0.19µV, with much less random jumps.

Some remain, but that might be caused by the BC337 which is 41 years old.
It could well suffer from popcorn noise.

This regulator should be able do as well as the nonoiser: 0.12 or 0.15µV
 
I have made a quick and informal check of the differential tempco of an ordinary Si transistor and an ordinary blue LED at comparable currents.

The match looks relatively good: the voltage difference is a shade over 2V, and for a ~75°C variation, the change is under ~40mV (positive).
Thus, the reference tempco is ~0.5mV/°C, not bad for a 2V reference.
It cannot match a true bandgap performance obviously, but for audio purposes, it is quite sufficient.

The advantage of using a blue LED is that the closed-loop gain is lower for a given output voltage, meaning better static performances, and better dynamic performances when no "cheat" caps are used.
One thing remains to be checked in this case: that the blue LED is not noisier than its green or yellow counterpart.
I'll give it a try one of these days.

Another advantage of the blue LED is the increased headroom for driving the small PMOS: even with a generic BS250 or similar, the regulator will become semi-LDO: about 1*Vbe, intermediary between the 2V or so of a traditional VR and the ~200mV Vcesat of a LDO.
To benefit from the increased headroom, the ratio between R5 and R9 should be altered.
 
Status
This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.