Low ripple low noise small power supply design

[gnuarm]

few updates:

- I have tried another new board, but same result happens. Also, I noticed that the 3.3v mosfet is really hot even at no load.

- using npn didn't give any good results, keeps being 5v but when putting load it drops and magic smoke happens due to the large voltage gap.

- I tried using 1.6 ohms and the result is 3.26v or so on 3.3v rail and 4.928v on 5v rail with no buzzing\continuous_beeping sound. can't go much because no heatsink installed.

- I have attached a picture of my wiring + pictures of real wiring (no shorting between any pin).


I started suspecting that the switchers + chosen KIA mosfets gonna make some change since that is the full circuit. And at the same time I find it bizarre that the circuit has some issues even with injecting voltages this way. having the full circuit allows us to get enough current on both rails without any drop voltage. do you think this might be the solution?

I don't understand your posts.  Here you show 12V directly connected to the drain of the MOSFET and yet you previously said the voltage was around 5V.  Did the previous measurements have a 5V regulator in the path to the drain? 

So what are the measurements in this configuration?   Whichever configuration you wish to trouble shoot let's stick with it and not jump around confusing the matter. 

No load test  of 3.3v rail:

v_gate = 1.35v
v_drain =4.64v from external buck module using same 12v source.
v_output_3.3v = 3.436v
v_opamp_supply = 12.38 supplied directly from 12v supply.
v_opamp_feedback = 3.446v
7805_5v_voltage_ref = 5.044v
7805_3.3v_voltage_ref = 3.360v, see note below.
v_12v_input = 12.38v.

I expect to see the two inputs to the op amp at virtually the same voltage.  The spec on the op amp is for low single digit mV input offset, so a difference of 14 mV should not be seen.  You also have 10 mV across the 3.3K feedback resistor which indicates 3 uA of bias current which is far above the low nA spec of the LM358. 

Any chance your op amp is a counterfeit part?  Did you buy them from a reputable source?

[BrianHG]

few updates:

- I have tried another new board, but same result happens. Also, I noticed that the 3.3v mosfet is really hot even at no load.

- using npn didn't give any good results, keeps being 5v but when putting load it drops and magic smoke happens due to the large voltage gap.

- I tried using 1.6 ohms and the result is 3.26v or so on 3.3v rail and 4.928v on 5v rail with no buzzing\continuous_beeping sound. can't go much because no heatsink installed.

- I have attached a picture of my wiring + pictures of real wiring (no shorting between any pin).


I started suspecting that the switchers + chosen KIA mosfets gonna make some change since that is the full circuit. And at the same time I find it bizarre that the circuit has some issues even with injecting voltages this way. having the full circuit allows us to get enough current on both rails without any drop voltage. do you think this might be the solution?

I don't understand your posts.  Here you show 12V directly connected to the drain of the MOSFET and yet you previously said the voltage was around 5V.  Did the previous measurements have a 5V regulator in the path to the drain? 

So what are the measurements in this configuration?   Whichever configuration you wish to trouble shoot let's stick with it and not jump around confusing the matter. 

No load test  of 3.3v rail:

v_gate = 1.35v
v_drain =4.64v from external buck module using same 12v source.
v_output_3.3v = 3.436v
v_opamp_supply = 12.38 supplied directly from 12v supply.
v_opamp_feedback = 3.446v
7805_5v_voltage_ref = 5.044v
7805_3.3v_voltage_ref = 3.360v, see note below.
v_12v_input = 12.38v.

I expect to see the two inputs to the op amp at virtually the same voltage.  The spec on the op amp is for low single digit mV input offset, so a difference of 14 mV should not be seen.  You also have 10 mV across the 3.3K feedback resistor which indicates 3 uA of bias current which is far above the low nA spec of the LM358. 

Any chance your op amp is a counterfeit part?  Did you buy them from a reputable source?

I was getting that feeling as well.

He needs a scope.  His volt meter readings may be all incorrect if he is measuring some sort of saw-tooth oscillation on those lines, but, not at the 3.3v reference point.

Without load, his 3.3v mosfet's output (source) should only have a 33ma load to GND, should it be getting hot?

If it is a nasty saw waveform, he might need to remove C38 & C39.

[VEGETA]

few updates:

- I have tried another new board, but same result happens. Also, I noticed that the 3.3v mosfet is really hot even at no load.

Hun?
Please measure the impedance of the mosfet pads to GND on the PCB.  No mosfet.  You should see 100 ohm, our current default pull-down resistor.  Definitely not enough to make a to220 mosfet get 'hot'.


Again, those 'Gate' voltages sound wrong.  A 1.35v from gate to GND should not output 3.436v, if anything, it should be close to 0v.

I measured the resistance of mosfet drain pad to ground (with no mosfet installed), result is 50.7k ohm. those gate voltages happen when no load, but when load is put they become reasonable.

I don't understand your posts.  Here you show 12V directly connected to the drain of the MOSFET and yet you previously said the voltage was around 5V.  Did the previous measurements have a 5V regulator in the path to the drain?

So what are the measurements in this configuration?   Whichever configuration you wish to trouble shoot let's stick with it and not jump around confusing the matter.

here is the one configuration I stick with:

No load test  of 3.3v rail:

v_gate = 1.35v
v_drain =4.64v from external buck module using same 12v source.
v_output_3.3v = 3.436v
v_opamp_supply = 12.38 supplied directly from 12v supply.
v_opamp_feedback = 3.446v
7805_5v_voltage_ref = 5.044v
7805_3.3v_voltage_ref = 3.360v, see note below.
v_12v_input = 12.38v.


NOTE: I noticed that before connecting power to mosfet drain, the 3.3v reference voltage which is derived from the resistor divider is only about 1.4v while the 7805 voltage is 5.044v. when connecting the mosfet drain power, it becomes 3.3v.


3.3 ohm load test  of 3.3v rail:

v_gate = 6.77v
v_drain =3.9v from external buck module using same 12v source. maybe it drops due to higher current
v_output_3.3v = 3.28v
v_opamp_supply = 12.11 supplied directly from 12v supply.
v_opamp_feedback = 3.337v
7805_5v_voltage_ref = 5.044v
7805_3.3v_voltage_ref = 3.331v, see note below.
v_12v_input = 12.11v.
buzzing completely gone

however, the post you quoted was another test after that, so let's stick with the ones written here.

I expect to see the two inputs to the op amp at virtually the same voltage.  The spec on the op amp is for low single digit mV input offset, so a difference of 14 mV should not be seen.  You also have 10 mV across the 3.3K feedback resistor which indicates 3 uA of bias current which is far above the low nA spec of the LM358.

Any chance your op amp is a counterfeit part?  Did you buy them from a reputable source?

well, lm358 bought from JLCPCB and LCSC, Chinese parts. I have previously bought ones but also from the same source. no way I can get ones now.

voltages on input pins of the opamp are different only when no load, but when putting load it gets nearly the same as shown in measurements... only few mV of difference.

the problem is that all 125 boards have the exact same parts, i could try to remove one opamp and install one of my older ones to see if it makes a difference.

[VEGETA]

I tried removing the opamp and install another one but failed, destroyed the footprint. so small i couldn't work with it.

[VEGETA]

On a side trial different than the one mentioned above, I removed the opamp by destroying it with heat... then soldered a buck module directly at the gate of the mosfet. now I outputted 4.6v to get 3.3v but it won't regulate properly on high loads.

so can we assume the fault is the opamp? since now no ringing or sound. what is the solution? I don't have the proper tools to remove such very small footprints as I tried now and couldn't remove it without brute forcing it. I paid a lot for the panels only to see it fail for no obvious reason... I am really frustrated now. Don't know what next step to take.

on a side note, a guy promised me to sell me Hanmatek DOS1102 digital oscilloscope for about 115$ after sending it to repair... he disappeared ffs!!

[gnuarm]

Do me one favor... you have tag lines linking to some web sites.  Add a link to someplace where I can see your schematic.  I keep having to dig back through the thread to find it.  As an image file it opens in MS photos which craps out after being open a while. 

[gnuarm]

few updates:

- I have tried another new board, but same result happens. Also, I noticed that the 3.3v mosfet is really hot even at no load.

Hun?
Please measure the impedance of the mosfet pads to GND on the PCB.  No mosfet.  You should see 100 ohm, our current default pull-down resistor.  Definitely not enough to make a to220 mosfet get 'hot'.


Again, those 'Gate' voltages sound wrong.  A 1.35v from gate to GND should not output 3.436v, if anything, it should be close to 0v.

I measured the resistance of mosfet drain pad to ground (with no mosfet installed), result is 50.7k ohm. those gate voltages happen when no load, but when load is put they become reasonable.

I don't understand your posts.  Here you show 12V directly connected to the drain of the MOSFET and yet you previously said the voltage was around 5V.  Did the previous measurements have a 5V regulator in the path to the drain?

So what are the measurements in this configuration?   Whichever configuration you wish to trouble shoot let's stick with it and not jump around confusing the matter.

here is the one configuration I stick with:

No load test  of 3.3v rail:

v_gate = 1.35v
v_drain =4.64v from external buck module using same 12v source.
v_output_3.3v = 3.436v
v_opamp_supply = 12.38 supplied directly from 12v supply.
v_opamp_feedback = 3.446v
7805_5v_voltage_ref = 5.044v
7805_3.3v_voltage_ref = 3.360v, see note below.
v_12v_input = 12.38v.

It would help if you included pin numbers so we know we aren't working around a communication error, but this is ok for now. 

NOTE: I noticed that before connecting power to mosfet drain, the 3.3v reference voltage which is derived from the resistor divider is only about 1.4v while the 7805 voltage is 5.044v. when connecting the mosfet drain power, it becomes 3.3v.

When you have no power on the MOSFET the output is a load on the feedback circuit.  The diode D3 connected that load to the V_3.3 signal which is a resistor divider and so is subject to change in value with loading.  When the FET is powered the circuit keeps the output very near the reference and the diode does not conduct.  I'm not sure the diode is required, but it should not hurt.  This is why V_3.3 is dragged down with no power on the FET drain or you will see the same thing if the op amp doesn't drive the FET. 

I was having a brain cramp.  I was reading gate voltage at 1.35V and thinking it was a bit low, but ok, then I realized that's not with a source at ground but at 3.4V!!!  So this is impossible.  The FET would be fully turned off and the source pulled down to ground by the 100 ohm R22. 

It is hard to think of a way you could have 1.2V between the source and drain when the gate is pulled so low.  if the source and drain were swapped you would see the body diode turn on, but the voltage between them would be well less than 1V with 33 mA of current.  Still, this almost has to be a wiring error with the FET or it is not the right part or a counterfeit part. 

I see you got parts from LCSC.  I don't know if their parts are real or counterfeit.  They sell Xilinx parts at less than a quarter of other distributors.  How can they do that if they are buying them from Xilinx?

3.3 ohm load test  of 3.3v rail:

v_gate = 6.77v
v_drain =3.9v from external buck module using same 12v source. maybe it drops due to higher current
v_output_3.3v = 3.28v
v_opamp_supply = 12.11 supplied directly from 12v supply.
v_opamp_feedback = 3.337v
7805_5v_voltage_ref = 5.044v
7805_3.3v_voltage_ref = 3.331v, see note below.
v_12v_input = 12.11v.
buzzing completely gone

however, the post you quoted was another test after that, so let's stick with the ones written here.

Ok, a drop in Vd of over 0.7V with 1 amp of current from a 2 amp supply is pretty poor performance.  With over 3V Vgs and 0.6V Vds the FET should be passing 15 amps.  Again, something is very wrong with the FET or your load resistor is not what you think it is.

I expect to see the two inputs to the op amp at virtually the same voltage.  The spec on the op amp is for low single digit mV input offset, so a difference of 14 mV should not be seen.  You also have 10 mV across the 3.3K feedback resistor which indicates 3 uA of bias current which is far above the low nA spec of the LM358.

Any chance your op amp is a counterfeit part?  Did you buy them from a reputable source?

well, lm358 bought from JLCPCB and LCSC, Chinese parts. I have previously bought ones but also from the same source. no way I can get ones now.

voltages on input pins of the opamp are different only when no load, but when putting load it gets nearly the same as shown in measurements... only few mV of difference.

the problem is that all 125 boards have the exact same parts, i could try to remove one opamp and install one of my older ones to see if it makes a difference.
[/quote]

You had 125 boards made before you debugged the circuit???  Wow! 

Vfb is about the same as Vref, but 6 mV offset is still outside the op amp spec even if only a little.  However the feedback is 60 mV off from the 3.3V output voltage so 18 uA into the -input on the op amp, well out of spec, about 1000x out of spec.  The diode will pass a tiny current, but much smaller than this when the voltage is only a few mV. 

To be sure of what you are measuring, please make the measurements directly on the pins of the op amps and the FET pins.  I see I was looking at the wrong op amp circuit previously.  This op amp circuit is pins 5, 6, 7.  The FET is still pins 1, 2, 3, or it might be more clear to call them G, D and S respectively.  Go back to your data sheets to make sure there's no mistake in the pin out.

Clearly something is wrong and you need to rule out every assumption possible.  Debugging is like being a detective.  Collect evidence and figure out what info is right and what info is wrong.  Someone in the circuit is lying to you.  Who can it possibly be?  Right now the FET looks to be the guilty party.  Try to verify the wiring and pin numbers. 

[VEGETA]

It would help if you included pin numbers so we know we aren't working around a communication error, but this is ok for now.

I sent the pdf schematic to you via PM.

Still, this almost has to be a wiring error with the FET or it is not the right part or a counterfeit part.

I made sure to check the wiring many times using the datasheet, and it is correct. The part is purchased long time ago from aliexpress but I think it is good since I used it in another electronic load circuit and worked as intended.

so according to schematic and layout diagram, pin 1 is gate on pin 1 of the foot print (left pin), pin 2 is drain which is the pad of footprint, pin 3 is the source which is pin 3 of the footprint (right pin).

I see you got parts from LCSC.  I don't know if their parts are real or counterfeit.  They sell Xilinx parts at less than a quarter of other distributors.  How can they do that if they are buying them from Xilinx?

I don't know about Xilinx parts but general parts like mosfets and npns are manufactured in China and won't be hard for them to make it. we based the design on this mosfet: https://lcsc.com/product-detail/MOSFET_KIA-Semicon-Tech-KND3403A_C382143.html which is a Chinese brand.

as you see, the same pin numbering between this and the one I tried.

Ok, a drop in Vd of over 0.7V with 1 amp of current from a 2 amp supply is pretty poor performance.  With over 3V Vgs and 0.6V Vds the FET should be passing 15 amps.  Again, something is very wrong with the FET or your load resistor is not what you think it is.

I have IRL640A mosfet to try, although I guess it gave me the same result but that was an old trial without high loading. I could try installing it and test if it makes sense but I don't think it is much different than iril44z mosfet.

You had 125 boards made before you debugged the circuit???  Wow!

I admit it is a mistake but we were kinda sure and optimistic about the design since no reason why it would fail or have a problem. I am still waiting for the mosfets and switchers hoping that it would make the circuit work. do you expect it would make a difference?

To be sure of what you are measuring, please make the measurements directly on the pins of the op amps and the FET pins.

I measured on the opamp pins and output connector. today I will try measuring the fet pins.

Clearly something is wrong and you need to rule out every assumption possible.

I am ready to do anything if it would make the boards work as intended. that is why when all parts arrive and get soldered, it will be the best test to do.

[gnuarm]

I don't know what to say.  I don't think adding parts will make it work better and I'm pretty sure it will make it harder to debug.  I think you should get rid of the 12 to 5V switcher you have now.  Run 12V to the FET.  I assume that supply will not droop with 1 amp of current, right?  Or test with a half amp, or a quarter amp if the FET gets too hot. 

Your circuit is not right at idle, so no need to load it really. 

I think you need to test Brian's idea that you might have oscillations.

[VEGETA]

I don't know what to say.  I don't think adding parts will make it work better and I'm pretty sure it will make it harder to debug.  I think you should get rid of the 12 to 5V switcher you have now.  Run 12V to the FET.  I assume that supply will not droop with 1 amp of current, right?  Or test with a half amp, or a quarter amp if the FET gets too hot. 

Your circuit is not right at idle, so no need to load it really. 

I think you need to test Brian's idea that you might have oscillations.

If I have oscillation, what to do to make the current boards work?

I am working hard to get a scope but the guy told its repair is not finished yet or no news from repair technician.

as for buck regulator, I ditched it in the last test as seen in the voltages.

today I couldn't do anything since I was tired but tomorrow I will for sure. my current plan is to use the board which I destroyed the on-board opamp... then use another opamp in its place by wiring it to the points in the board to see if lm358 is the reason.

brian said stuff about removing 1nf caps... will this really work? I mean do it affect oscillation being so little? I think it slows the reaction a little bit but make the entire circuit bad?? also, the sound issue is still not known. removing these caps is very easy and if it will fix the circuit then that would be a happy solution.

[gnuarm]

I don't know what to say.  I don't think adding parts will make it work better and I'm pretty sure it will make it harder to debug.  I think you should get rid of the 12 to 5V switcher you have now.  Run 12V to the FET.  I assume that supply will not droop with 1 amp of current, right?  Or test with a half amp, or a quarter amp if the FET gets too hot. 

Your circuit is not right at idle, so no need to load it really. 

I think you need to test Brian's idea that you might have oscillations.

If I have oscillation, what to do to make the current boards work?

Figure out why it is happening first, that will tell you how to fix it. 

I am working hard to get a scope but the guy told its repair is not finished yet or no news from repair technician.

as for buck regulator, I ditched it in the last test as seen in the voltages.

Sorry, you've lost me.  The last numbers I see show

v_drain =4.64v from external buck module using same 12v source.

How could the drain be 4.64V if it is connected directly to 12V without the switcher?

today I couldn't do anything since I was tired but tomorrow I will for sure. my current plan is to use the board which I destroyed the on-board opamp... then use another opamp in its place by wiring it to the points in the board to see if lm358 is the reason.

I don't know what to say.  There are so many mistakes made because you thought you could get this to work on the first pass without any debug.  Using small pitch surface mount parts when you don't have the means of soldering them is a big mistake.  Making 125 boards as your first run was a huge mistake.  I am floored that you attempted this.  I can easily afford such a mistake but I will always make a prototype before I run a production batch. 

brian said stuff about removing 1nf caps... will this really work? I mean do it affect oscillation being so little? I think it slows the reaction a little bit but make the entire circuit bad?? also, the sound issue is still not known. removing these caps is very easy and if it will fix the circuit then that would be a happy solution.

I don't know why you think 1 nF is "little".  A few pF in the wrong place can cause oscillations with a fast circuit or a high gain circuit.  The 1 nF cap is doing nothing useful in this design anyway.  So remove it.  It was only there as part of a "capacitance multiplier" concept that doesn't apply to this circuit because it is a regulated output.  The added capacitance can only slow the response of the feedback path which causes oscillations.

[VEGETA]

Sorry, you've lost me.  The last numbers I see show

12v directly and buck module gave the same results. i am sorry if i didn't document the 12v directly one, it will be my next thing to do soon.

I don't know what to say.  There are so many mistakes made because you thought you could get this to work on the first pass without any debug.  Using small pitch surface mount parts when you don't have the means of soldering them is a big mistake.  Making 125 boards as your first run was a huge mistake.  I am floored that you attempted this.  I can easily afford such a mistake but I will always make a prototype before I run a production batch. 

well, I told you we were confident that it will work since nothing in the circuit is weird or special circuit. still, it is a mistake to do and I should learn from it.

The 1 nF cap is doing nothing useful in this design anyway.  So remove it.

I will remove it for sure. gonna keep u updated with what happens.

[gnuarm]

Sorry, you've lost me.  The last numbers I see show

12v directly and buck module gave the same results. i am sorry if i didn't document the 12v directly one, it will be my next thing to do soon.

You may think it is the same results, but the Vd will be different.  Every value you measure is important in the context of the others.  Change Vd and Vds changes which presents a very different picture of what is going on.  If Vd doesn't change there is something very, very wrong.

I don't know what to say.  There are so many mistakes made because you thought you could get this to work on the first pass without any debug.  Using small pitch surface mount parts when you don't have the means of soldering them is a big mistake.  Making 125 boards as your first run was a huge mistake.  I am floored that you attempted this.  I can easily afford such a mistake but I will always make a prototype before I run a production batch. 

well, I told you we were confident that it will work since nothing in the circuit is weird or special circuit. still, it is a mistake to do and I should learn from it.

Not trying to beat you up about it.  Yes, a learning experience.

The 1 nF cap is doing nothing useful in this design anyway.  So remove it.

I will remove it for sure. gonna keep u updated with what happens.

Ok

[VEGETA]

1nf caps turned off + no load on both rails:

power:
v_opamp_supply = 12.36
v_12v_input = 12.36v.

3.3v rail:
v_gate = 1.277v
v_drain = 12.36v
v_output_3.3v = 3.432v on connector + 3.432v on pin.
v_opamp_feedback = 3.442v
7805_3.3v_voltage_ref = 3.360v


5v rail:
v_gate = 1.598v
v_drain = 12.36v
v_output_5v = 5.116v
7805_5v_voltage_ref = 5.041v
v_opamp_feedback = 5.126v


same sound exist.

however, both now and before removing caps... i noticed when i press or touch R17  sound gets... faster.

[VEGETA]

1.6 ohm load:

power:
v_opamp_supply = 11.02v
v_12v_input = 11.06v.

3.3v rail:
v_gate = starts at 6.27 and keeps rapidly decaying, could not keep it up due to heat but it went below 5.7 and was still decaying. the solder melted while doing it due to heat and mosfet fell xD
v_drain = 10.65v
v_output_3.3v = 3.204v on connector + 3.255v on pin. it increases slowly but could not keep it up due to heat.
v_opamp_feedback = 3.289v
7805_3.3v_voltage_ref = 3.271v

[gnuarm]

The numbers on your circuit are all wrong and I can't tell you how to fix it without an oscilloscope to see what is going on.  You must have something oscillating. 

One thought.  It is very odd that the gate voltage is below the source voltage.  Try connecting the op amp output to the - input (V_FB1) short term.  The 3.3K R19 will allow the op amp output to control the input.  Then the gate voltage should be 3.3V regardless of load.  I'd like to see what the output is then. 

For the short term if I were you, I would focus on getting the switching regulators working and use this for the 3.3 volt output and a similar part for the 5V output.

https://www.digikey.com/en/products/detail/stmicroelectronics/LD1085D2M33R/669198

This will at least get you a working circuit to start with.  Then you can try some things with the FET circuit. 

[VEGETA]

if I connect pin 7 which is output of opamp to R19 (or the 3.3v output) it gives about 53mv on the 3.3v output while using 1.6 ohms load.

getting such LDO is good if I want to rework the circuit entirely since it won't fit on the current board. I don't think they will fit in the same board and layout.

If I am gonna re-do the circuit, I would probably use a good switcher IC with very little noise such as TPS62913 but this one still not active part, very new. Another part is LM61460, this one can deliver good noise\ripple (<10mV) but can't beat TPS62913 which can get below 1mv despite being a switcher. these ICs are not cheap but if I could get them to replace the linear stage + heatsink + extra components (opamp+7805) then I go for it


actually the original design of this psu uses similar approach to yours. see it in attachment. what do you think? it was very straight forward but I wanted cap multi and then you know what happened.

however, i still want to think there is a way to salvage these boards if we could solve the problem.

[Vovk_Z]

- I have attached a picture of my wiring + pictures of real wiring (no shorting between any pin).

I think you using C38, C38 1nF caps wrong. Remove them from gates, and put them between 1-2 and 67 legs of LM358.
...
And you may try lower R16, R17 a bit (22-47 R). But this is not the main problem.

[BrianHG]

I do not know how you are feeding the drains on the mosfets, but, I would wait until you install the AOZ1284 switcher ICs and power the PCB the way it was designed to be powered.  You cannot Gerry rig in a third party switcher which may be under-powered and inject ground bounce along long thin wires and connection to your uncertain

As for C38 and C39, yes, those should not be installed.

After the 2 AOZ1284 are installed, if there is oscillation because of a cheap op-amp, the easiest way to control it would be to place a cap 1nf, 10nf, 100nf cap between the mosfet gate and source.

I am not sure why your mosfets are conducting/feeding 3.3v out of the source when there is only 1.25v on the gate.  Are you sure your volt meter's GND is connected to the PCBs true GND?

[gnuarm]

if I connect pin 7 which is output of opamp to R19 (or the 3.3v output) it gives about 53mv on the 3.3v output while using 1.6 ohms load.

What is the voltage on pin 7 of the op amp?  Is that still about 3.3V?  Connecting to the +3V3 is a bad idea.  The point is the gate to drain voltage is wrong in all cases.  It is far too low with no load. 

Oh, and try testing at a current level that doesn't melt the solder!!!

getting such LDO is good if I want to rework the circuit entirely since it won't fit on the current board. I don't think they will fit in the same board and layout.

I'm pretty sure you can make it work the way you have things wired up now.  If the regulator doesn't work then maybe there is something wrong with the circuit board.

If I am gonna re-do the circuit, I would probably use a good switcher IC with very little noise such as TPS62913 but this one still not active part, very new. Another part is LM61460, this one can deliver good noise\ripple (<10mV) but can't beat TPS62913 which can get below 1mv despite being a switcher. these ICs are not cheap but if I could get them to replace the linear stage + heatsink + extra components (opamp+7805) then I go for it

This is why I asked what your goals/requirements were.  Do you really need to get ripple down to 1 mV?  You are trying to make it work in an application, what does that application require?

actually the original design of this psu uses similar approach to yours. see it in attachment. what do you think? it was very straight forward but I wanted cap multi and then you know what happened.

however, i still want to think there is a way to salvage these boards if we could solve the problem.

This is why I asked about the fixed gate drive.  You can use an NPN with a bias to V3.3 +  a diode drop to drive the base.  Then the gain of the transistor will smooth out the ripple. 

But why would you think any other circuit won't have a similar problem to this circuit?  Something is going on that you are not aware of.  This is a learning experience.  It's also an opportunity to learn how to use a scope.

What is the speed of the switching on the buck converters?  Certainly not over 1 MHz or not by much anyway.  You can get a 20 MHz digital USB scope to use with your computer for around $100.  That should work with your problem to let you see what is going on.  There are some open threads here for low end units.

[gnuarm]

I do not know how you are feeding the drains on the mosfets, but, I would wait until you install the AOZ1284 switcher ICs and power the PCB the way it was designed to be powered.  You cannot Gerry rig in a third party switcher which may be under-powered and inject ground bounce along long thin wires and connection to your uncertain

As for C38 and C39, yes, those should not be installed.

After the 2 AOZ1284 are installed, if there is oscillation because of a cheap op-amp, the easiest way to control it would be to place a cap 1nf, 10nf, 100nf cap between the mosfet gate and source.

I am not sure why your mosfets are conducting/feeding 3.3v out of the source when there is only 1.25v on the gate.  Are you sure your volt meter's GND is connected to the PCBs true GND?

Perhaps it would be better to measure each important parameter point to point, Vgs, Vds, etc? 

I have no idea what the long wires would be doing to a circuit like this.  The op amp is only about 1 MHz GBW i seem to recall, so a bit of stray capacitance shouldn't be that important, but still! 

[BrianHG]

Perhaps it would be better to measure each important parameter point to point, Vgs, Vds, etc? 

I have no idea what the long wires would be doing to a circuit like this.  The op amp is only about 1 MHz GBW i seem to recall, so a bit of stray capacitance shouldn't be that important, but still!

Yes.
Though, as the mosfet heats up, the voltage on the gate will decay since that figure is temperature dependent.
The opamp seems to have a huge differential offset.  Probably a fake LM358N.  This is not the quality I would like, but they can be still used if they don't oscillate, or worst case, a cap may be need across the mosfet gate and source to feedback/null the high frequency.

Without a scope, we cannot tell where that buzzing is coming from.  It might have something to do with his source switcher having a low power mode which pulses the power out.  Sending 12v to the mosfet drains will heat them up FAST without a heat sink.  They will still heat up without any heatsink even if you feed them 4.3v.  The method of sending 4.3v with an unknown switcher was never foreseen as a design consideration for this PCB.  The power loop and GND on that PCB was designed and is expected to be tight where the AOZ1284 switchers feeding the drains were designed to go from 0 amp load to 4 amp load, in constant frequency mode.

[gnuarm]

Testing does not need to be done at 2 amps.  The 100 ohm built in load is good enough.  Add a 10 ohm load to see how it does with a bit more current.  Measurements can be taken before the solder melts.  Provide power from a bench supply.  Add a resistor between the 12V rail and the drain to dissipate power.  There are a million ways to manage that. 

I don't know what to do here.  The op amp does not behave 100% correctly, the FET is really anomalous passing current with the gate at negative voltages wrt the source. 

If I were debugging the board I would be looking for something mechanically wrong like bad soldering or wrong connections.  I guess the board is being used because the small parts are on it.  I take it the FET is not the part designed in and so the wires to connect it?  I don't even see a part number on the schematic. 

If the switchers are to be mounted and tested, then the FET needs to be left off until the switcher is properly tested. 

He tried driving the gate from a fixed 3.3V, maybe drive it from a fixed 5V, but not a 2 amp load!  There is something wrong with the FET and it needs to be figured out. 

Heck attach a pot to the 12V supply with the wiper tied to the FET gate.  Take output measurements at every setting between 0 and say 4.5V, but with a 10 ohm load maybe. 

[VEGETA]

I will try putting a potentiometer at the gate of the mosfet to give its gate voltage from 0-12v. I hope connected opamp won't affect it. output load will be 47R for both rail.

also, I will try putting ceramic caps on mosfet source to gate.

switchers and original mosfets will need more time to arrive.

[gnuarm]

I will try putting a potentiometer at the gate of the mosfet to give its gate voltage from 0-12v. I hope connected opamp won't affect it. output load will be 47R for both rail.

also, I will try putting ceramic caps on mosfet source to gate.

switchers and original mosfets will need more time to arrive.

Don't do all these things at one time.  The op amp needs to be disconnected when you attach the pot wiper to the gate.  You can lift R17.  What FET are you using?  I've been looking at a data sheet for IRLIZ44NPbF.  Still, no FET should be on with a negative Vgs unless it is enhancement mode and I don't think they make those in power versions.

When you do soldering on the PCB you need to do it quickly.  Holding the iron on the board for extended times is not good for the tracks.