Multi-phase buck converter circuit — strange behavior.

[Komajster]

Hello!
I have a broken PC graphics card (EVGA GTX 970 SSC) with 4-phase GPU vcore power, one day the two low-side FETs from one of the buck converters just decided to short themselves and weld themselves to the copper on the PCB creating a permanent short circuit inside it. I was able to remove this short and make the card work with remaining 3 buck converters making power for the GPU ASIC. Now I need to restore the 4th phase, I recreated the destroyed traces on and inside the board, soldered down a new set of FETs (there are two low-side and one high-side), as well as a new driver for them (NCP81062). I did check all the connections based on the application circuits from driver and the buck controller datasheets (https://www.onsemi.com/pub/Collateral/NCP81062-D.PDF and https://www.onsemi.com/pub/Collateral/NCP81174-D.PDF), but now when I power on the card the controller seems to enable all the four converters for a couple milliseconds and then it disables them. Looking at the waveform of each phase (at the switch plane, before the output coil) I can see the probable reason for that, this is how it looks like on the 3 working phases:

(as expected, 300 kHz PWM signal)

And on the 4th phase:

And this doesn't look like it should do...
I don't know how to interpret that signal, maybe there will be someone that could tell me something more about what it could mean.

All of the NCP81174's PWM outputs look the same:

(300 kHz PWM signal)

At this point I don't know why it doesn't work, I already tried three different NCP81062 drivers from different batches, with the same result every time, I swapped the main buck controller (NCP81174), changed every FET thinking that they can be the issue, I even tried with a different output coil, but nothing seems to change anything in the way it behaves now.

Here is how the reconstructed buck converter and the whole power delivery section look like: https://imgur.com/a/eHyJdEr

[T3sl4co1l]

Is the ground side via'd to plane?

Tim

[Komajster]

Is the ground side via'd to plane?

You meant this part of the application schematic?



Then yes, it is, the R3 and C6 are just swapped and the C6 is 1 nF and R3 is 1 Ω.

Here is how it looks like, you can see the gray wire from the other side of the board (as shown at the end of my first post) is connected here to the cap:

[T3sl4co1l]

I mean these:

That gray wire is looping all the way around?  Like, I guess there was a via originally, but it got disintegrated in the fire..?  Yeah that's not doing a damn thing.  Loose wire has an inductance around 0.6 uH/m, and I'd guess that has a length around 0.1m so an inductance of ~60nH.  The RLC circuit thus has a resonant impedance of sqrt(L/C) = sqrt( (60nH) / (2.7nF) ) = 4.7 ohms, higher than the 2.2 ohm resistance (giving it a Q around 2), and a resonant frequency of 12.5MHz.

The purpose of that network, is to dampen either the resonance of the switching loop (the stray inductance between the two transistors and the supply plane), or the free ringdown of the inductor (which occurs at light load, when current falls to zero and both transistors turn off).  For the former purpose, the stray inductance in the RC path needs to be less than the switching loop, i.e. a few nH.  Which they probably didn't achieve in the original layout, by the look of it.  But also the R required for damping the ringdown should be larger (around sqrt(235nH / 1.5nF) = 12.5 ohms), so I wonder what they were even after.

Tim

[Komajster]

I mean these:

These are connected right into the PCIe power connector ground via the thick black wire.

Like, I guess there was a via originally, but it got disintegrated in the fire..?

Yes, that is the case here.

The purpose of that network, is to dampen either the resonance of the switching loop (the stray inductance between the two transistors and the supply plane), or the free ringdown of the inductor (which occurs at light load, when current falls to zero and both transistors turn off).  For the former purpose, the stray inductance in the RC path needs to be less than the switching loop, i.e. a few nH.

Thank you for this explanation, I moved the network to the top side to get rid of that loose wire, I did put it physically between the two low-side FETs (is this ok?). Now the resonant frequency measured with the scope seems to be lower...? But I think that it should be higher because the inductance is lower? It's about 6 MHz if I'm measuring it correctly.

And of course the buck controller still disables all the converters in a very brief time after power on.

[T3sl4co1l]

These are connected right into the PCIe power connector ground via the thick black wire.

Oh yikes.  Oh no, that won't do. That won't do at all.

If you must, dig into the board until you find the ground plane.  Make as wire as possible a connection to it.

This is rather heroic effort, and I guess you will appreciate why these things aren't usually repaired.

(The board can even be rebuilt by pouring epoxy in the pocket thus created, re-laying traces over and everything.  Needless to say this is only ever done with things that simply cannot be replaced -- ancient military gear, restorations.)

Thank you for this explanation, I moved the network to the top side to get rid of that loose wire, I did put it physically between the two low-side FETs (is this ok?). Now the resonant frequency measured with the scope seems to be lower...? But I think that it should be higher because the inductance is lower? It's about 6 MHz if I'm measuring it correctly.

Yes, that is better.  But because the bottom side transistors are already on, they don't do anything to this transient.

A similar network across the high side transistor can dampen this transient, but it can't do anything about the peak voltage.  The peak negative voltage is most likely what's stressing the gate driver / controller.  I'm impressed it didn't just explode after a few cycles of that waveform...

Tim

[Komajster]

If you must, dig into the board until you find the ground plane.  Make as wire as possible a connection to it.

This is rather heroic effort, and I guess you will appreciate why these things aren't usually repaired.

As heroic and not worth the time as it might be, I did it, I dug into the board, found the ground plane and made a strong connection, very close to where the original vias were. Now the waveform is identical to what I see on the other three phases, there isn't any major negative voltage peak, but the controller still does the same thing... I could probably damage it in some way by pushing the -17.4 V into one of its pins over and over again, should finding another controller be the next thing to try?

[T3sl4co1l]

  Well done sir!

Good question, then... I wonder if the controller got toasted from the abuse -- I would suspect it, yes.  A replacement would be worth a try.

Tim

[Karalux]

Hello everyone, i hope i wont disturb the purpose of this post but i am writing solely out of desperacy looking for help with this gpu (well, mine is SC model 2977-KR, but vrm circuits are nearly identical). I hope you could help me with the situation i have, which is that i tried to fix my gpu by replacing 4C10N mosfets and its NCP81062 controllers and accidentally blew away couple capacitors and lost them. I thought maybe you guys could help me to identify what value these components were so i could finally fix my gpu and use it (it turned on when i desoldered the broken modfet). All i need is the two caps in red circles and if possible other 4 components in small green circles too. Hope someone could help me since ive tried russian/ukrainian forums and even contacting EVGA service and Kingpin himself with no luck. Thank you in advance

[Komajster]

Hello everyone, i hope i wont disturb the purpose of this post but i am writing solely out of desperacy looking for help with this gpu (well, mine is SC model 2977-KR, but vrm circuits are nearly identical). I hope you could help me with the situation i have, which is that i tried to fix my gpu by replacing 4C10N mosfets and its NCP81062 controllers and accidentally blew away couple capacitors and lost them. I thought maybe you guys could help me to identify what value these components were so i could finally fix my gpu and use it (it turned on when i desoldered the broken modfet). All i need is the two caps in red circles and if possible other 4 components in small green circles too. Hope someone could help me since ive tried russian/ukrainian forums and even contacting EVGA service and Kingpin himself with no luck. Thank you in advance (Attachment Link)

Here you go.

[Karalux]

If its true then i dont even know how to thank you, i wish i could help with your problem. Did you actually measure the components or its from the controller specification? Ive seen the ON Semiconductors NCP81062 specification with suggested connection scheme and i was too afraid going just by that, since EVGA might use different values resistors and capacitors

[Komajster]

If its true then i dont even know how to thank you, i wish i could help with your problem. Did you actually measure the components or its from the controller specification? Ive seen the ON Semiconductors NCP81062 specification with suggested connection scheme and i was too afraid going just by that, since EVGA might use different values resistors and capacitors

I did measure them physically for one of the phases and then with the rest I just tracked down where the traces go on my card with identical VRM. I'm sure that the values I wrote there are correct.

[Karalux]

Thank you very much, ill try to solder everything back tomorrow and see for myself, hope it will finally work

[Karalux]

So i soldered everything back, then tried the gpu and all i got was just a pop and some "magic smoke". Something shorted out and after taking all the mosfets out i found out that same low side 4C10N is dead again, but this time one high side 4C06N is dead too. Is there any way to check NCP81062 controller if it is dead or not? I dont want to risk putting dead controllers back, since i only have one left which is still good and a bunch of others that some are good, some are probably not, i have to check them 

[Komajster]

So i soldered everything back, then tried the gpu and all i got was just a pop and some "magic smoke". Something shorted out and after taking all the mosfets out i found out that same low side 4C10N is dead again, but this time one high side 4C06N is dead too. Is there any way to check NCP81062 controller if it is dead or not? I dont want to risk putting dead controllers back, since i only have one left which is still good and a bunch of others that some are good, some are probably not, i have to check them 

As far as I know there isn't any easy way to definitely tell if the driver is dead or not, you can check the resistances from the gate driving pins (DRVH and DRVL) to chip's ground, it should be in kilohms (I just measured it on a working driver, it's about 67K), but that doesn't really tell you if the driver is working. I've had a couple of these cards and generally if any FET was shorted I also replaced the driver, they can very easily die after the mosfet shorts out since it can send some current to the gate driving pin and kill the driver. Sorry for a late reply but I have a lot of work now.

Going back to the original problem of this post...

Well done sir!

Good question, then... I wonder if the controller got toasted from the abuse -- I would suspect it, yes.  A replacement would be worth a try.

Thank you for the congratulations I am still waiting for the new controller to arrive.

[T3sl4co1l]

Usual suspicions are, measure node resistance or diode drop when unpowered.  May not be all that meaningful in-circuit, because of the big beefy supply rails that won't budge from the small current of a diode test.  But anyway, you expect most pins to measure a diode from GND to pin, and pin to whatever +V is for the chip (or that part of it), usually around 0.6V.  If it's tiny (less than 0.3V probably), it may be a resistance, which may be normal (also, check for connected pull-up/down resistors), or an indication of damage.

Tim

[boB]

If you are still working on the issue of the controller turning off soon after turning on, I would check the R-C current sense feedback to the controller for that one bad cylinder of the 4.

Maybe the resistor opened up ?