Mosfet paralleling reliability in real life - more smaller ones vs fewer bigger

[Miyuki]

Hi folks
I have a question about reliability of power supply with higher number of paralleled transistors

For example 2 Switch forward converter with planned 2x 5 in parallel switched at moderate frequency 200kHz
If current will be equally shared losses are calculated ~10W per device 4.5:5.5 Conductive:Switching losses ratio

Question is if then I can trust that this many devices have enough similar parameters to work reasonably and over longer time period
Yes then here is also some layout issue with switching parallel at this speed, but there little extra trace inductance at Drain side can be beneficial

[MagicSmoker]

Five MOSFETs in parallel doesn't strike me as terribly economical unless you are relying on passive convection cooling and need to get the heat flux (ie - W/cm²) down.

Paralleling MOSFETs isn't usually a big deal as long as you increasingly derate the current for each additional switch and use an individual gate resistor for each one, however at 200kHz you might have a problem with differing loop impedances which will lead to uneven dynamic current distribution. Adding a bit of inductive impedance in series with the source of each switch can be helpful here; putting the L in series with the drain is less effective because you don't get the benefit of negative feedback from degeneration.

That said, it is always preferable to minimize the number of switches in parallel, and (5) seems a bit much unless this is a step-up application and the input voltage is low (e.g. - 12V to 340VDC to supply the H-bridge in an inverter).

[Miyuki]

But five small is actually cheaper than one big and I can use fullpak ones with simpler and cheaper mounting
Plus I need to switch relative high current at short duty, about 20-30A it makes problem with single device and high switching losses

[rx8pilot]

This YT channel is gold.

https://youtu.be/wjSM2oYY3OY

[MagicSmoker]

But five small is actually cheaper than one big and I can use fullpak ones with simpler and cheaper mounting
Plus I need to switch relative high current at short duty, about 20-30A it makes problem with single device and high switching losses

30A isn't particularly challenging for a single device or, perhaps, two at most. What is the maximum supply voltage?

[Miyuki]

Supply voltage is 500-650V what calls for at least 800V transistors where choice is limited

[schmitt trigger]

In the end, like every thing else in engineering, there will be tradeoffs.

My pragmatic advice, is if this is a high volume, cost sensitive product with a proper design budget, then by all means optimize the design. This may mean that a few iterations may be required to satisfy your cost, performance and reliability goals.

In the other hand, if the design is for very few units and you are on a tight design timeline, go for the simplest approach. Use a minimum of larger Mosfets.

[MagicSmoker]

Supply voltage is 500-650V what calls for at least 800V transistors where choice is limited

Well that certainly changes things... You need to limit transition time to around 25ns to keep switching losses manageable but that results in something like 26V/ns dV/dt which is way too much for hard-switched operation, especially when body diode conduction is involved (which it will be in the two-switch forward). Also, 19.5kVA is a bit much for a two-switch forward... like 10x too much. Granted, there's not as much of a penalty in transformer utilization between a single ended and bridge converter at this frequency, but there still is one. At this power level (and input voltage range) the full bridge converter is pretty much exclusively used, a shaving a few $ off the switch cost is not even a consideration.

[Miyuki]

Transmited power is "just" about 1500W, but it is constant current / power and need high possible output voltage at low current
Why would body diode conduct any bigger portion of time, it is unwanted and just some nasty osculations do it.

With small parallel transistors current in each shall be enough small to allow reasonable switching speed and losses.

With one device I will need to go with SiC device or accept high losses and problem with cooling

[MagicSmoker]

...
Why would body diode conduct any bigger portion of time, it is unwanted and just some nasty osculations do it.

The energy stored in the transformer magnetizing and leakage inductances (and any other stray inductance in the loop) is returned to the supply via the body diodes each switching cycle. Even in fast 800V-900V SuperJunction MOSFETs these diodes will have a reverse recovery time in the >300ns range, which is way too slow for operation at 200kHz.

The body diodes in SiC MOSFETs are faster,  but they are PN type, not Schottky, so still have a reverse recovery time (30ns - which is fast enough - to 60ns - which is marginal) and a really high forward voltage drop (>4V in most cases) because of SiC's wide bandgap.

As a result, handling this much power in a single converter at 200kHz practically begs for a (quasi-)resonant topology.

[T3sl4co1l]

Why not a pair of IGBTs?  Should be just workable at that frequency, or fine a bit lower.  If you're going for lower switching losses in a compact package (hence the higher frequency) I guess that wouldn't work out; instead I would recommend SiC MOSFETs.

Paralleling MOSFETs is generally fine, but don't forget to consider stray inductance.  In the Sam Ben-Yaakov video, he considers the current divider circuit formed by Rds(on)s.  Add stray inductances in series with these, and change the current to a step or ramp source, and you will have a representation of current sharing in the switching circuit.  Current equalizes after a few L/R time constants, so besides reducing stray inductance, it can actually help to have some additional resistance.

Note that the stray inductance is, always, the loop inductance, from one switch to its opposing switch.  For the two-switch forward, that's one transistor, and its catch diode, through the local bypass capacitor.  The other transistor and diode act in an independent loop, and that loop needs to have the correct inductance as well.  (The bypass can be shared, that's fine -- as long as both inductances work out.)

A typical layout, with transistors lined up on a heatsink, you would need to alternate transistor and diode, and put bypass caps in front of the row.

It's not much more trouble to simply use fully independent inverters.  Use a single transistor and diode per leg, and a pair of legs, with a transformer and all the support components (input bypass cap, driver, output rectifier..) per channel or phase.  Use N phases, with a phase shift between each, to reach the desired total capacity.

This is a bit annoying to do at 1500W, so I don't know that I would bother to do it (probably, I would choose a single phase, or two phase interleaved, full wave, forward converter).  It is the only practical way to scale up arbitrarily.

The fundamental problem with scaling, is what ultimately underlies this argument.  It becomes much more apparent at low voltages: where stray inductances hit so much harder.  Say you're doing a 12V to 1.0V Vcore supply on a motherboard.  You need 100A out, so you're switching pulses of about 100A / N at the input, for N phases.  A hundred amperes in a single stage is just preposterous: that's a 0.12 ohm switching load impedance to begin with, and even doing it at 100kHz, you have to contend with the reactance of even a very compact loop of 5nH being 3mohm, i.e., drawing a reactive power of 2.6% of the total power, give or take.  And that's just at Fsw; basically all the harmonics are going into it as well.  That reactive power is just going to be burned as switching loss, unless you go to lengths to conserve it (quasi-resonant snubber?), and even then, you can't snub very much of it because the snubber itself is going to have about as much stray inductance!

So you need to divide and conquer, and 10A per phase is far more manageable, indeed well enough that Fsw can be pushing 1MHz while switching losses stay quite comfortable.

Finally, the other side is this: why not just build a bunch of inverters and run them in parallel?  Why phase interleave?  The best part is, when phases are interleaved, their ripple currents interfere and partially cancel out.  You can save on total capacitance and filtering this way.

Tim

[T3sl4co1l]

The energy stored in the transformer magnetizing and leakage inductances (and any other stray inductance in the loop) is returned to the supply via the body diodes each switching cycle. Even in fast 800V-900V SuperJunction MOSFETs these diodes will have a reverse recovery time in the >300ns range, which is way too slow for operation at 200kHz.

Reverse recovery is only relevant in hard switching (CCM).

Flyback is typically operated in DCM, so this doesn't matter (for certain degrees of "matter" -- in the two-switch, the diodes act in series and one will inevitably recover before the other, leaving some recovery charge left in the other one).

Although I suppose you wouldn't be doing DCM at this power level, but you shouldn't be forcing a flyback design at this power level, either.

On a related note, in ZVS (inductive load) configurations (including resonant), body diode recovery occurs during normal load current, while the transistor is on -- this probably lengthens the recovery time (because the recovery voltage drop is small?) but it means most MOSFETs are adequate to a MHz or more, even high voltage Si being adequate at half a MHz or more.

Tim

[MagicSmoker]

Reverse recovery is only relevant in hard switching (CCM).

Flyback is typically operated in DCM, so this doesn't matter (for certain degrees of "matter" -- in the two-switch, the diodes act in series and one will inevitably recover before the other, leaving some recovery charge left in the other one).

Err... the OP first said this is a two-switch forward, not a flyback.

...
On a related note, in ZVS (inductive load) configurations (including resonant), body diode recovery occurs during normal load current, while the transistor is on -- this probably lengthens the recovery time (because the recovery voltage drop is small?) but it means most MOSFETs are adequate to a MHz or more, even high voltage Si being adequate at half a MHz or more.

Well, someone hasn't had their coffee yet this morning...  Re-read my previous post. 

[T3sl4co1l]

Err... the OP first said this is a two-switch forward, not a flyback.

Note sure how I internalized it as flyback, but it works out the same on the primary side at least.

Transformer utilization is half, relative to full wave; I wouldn't call it insignificant at scale.  There may be many costs much greater than the transformer alone, but in an absolute sense, no.

Well, someone hasn't had their coffee yet this morning...  Re-read my previous post. 

Just reiterating it in case it hasn't become crystal clear to the OP

Tim

[MagicSmoker]

Note sure how I internalized it as flyback, but it works out the same on the primary side at least.

Transformer utilization is half, relative to full wave; I wouldn't call it insignificant at scale.  There may be many costs much greater than the transformer alone, but in an absolute sense, no.

Oh, I totally agree that the (hard-switched) two-switch forward is not even remotely the topology of choice here (despite immunity from cross-conduction failure, which is otherwise very compelling), but the penalty in power throughput at >100kHz is not nearly as bad as you might think because max flux swing has to be greatly curtailed to keep core losses under control, anyway. Practically speaking, the advantage in transformer utilization of bipolar vs. unipolar operation at fsw >100kHz is more like 20%, rather than a doubling. Still, one shouldn't be hard-switching 650V at 200kHz, anyway, so rather a moot point.

[T3sl4co1l]

the penalty in power throughput at >100kHz is not nearly as bad as you might think because max flux swing has to be greatly curtailed to keep core losses under control, anyway. Practically speaking, the advantage in transformer utilization of bipolar vs. unipolar operation at fsw >100kHz is more like 20%, rather than a doubling. Still, one shouldn't be hard-switching 650V at 200kHz, anyway, so rather a moot point.

Ah yeah -- that's one of those crusty old things that I've manged to internalize without consideration!  Thanks for bringing that up.

Back in the day down at like 20-50kHz, you'd be able to push Bsat, but it's still hard today to find materials that will handle that much at 200kHz (even 100kHz in modest sizes).  Modern materials like N97 and 3F46 are quite good, but not quite that good.  And that explains the relative prevalence of otherwise-kinda-shitty topologies -- like two-switch forward -- in modern application.

So put that on the bonfire of obsolete design information as well.

Tim

[Miyuki]

I would like to use ETD shape core, it have relative big area to volume ratio so I can push flux swing high (about 200mT at 200kHz) with sustainable temperature rise
But it is far far from saturation even at single ended topology

Also switching is just half hard, at turn off since turn on is covered by leakage inductance at almost zero current and have just Coss losses

//edit:
wonder what will do to switching losses driving mosfets off very hard with like -10V with minimal gate resistance and rely on diodes to clamp that spike

[mzzj]

Transmited power is "just" about 1500W, but it is constant current / power and need high possible output voltage at low current
Why would body diode conduct any bigger portion of time, it is unwanted and just some nasty osculations do it.

With small parallel transistors current in each shall be enough small to allow reasonable switching speed and losses.

With one device I will need to go with SiC device or accept high losses and problem with cooling

This sounds pretty much like welding machine..

15 years ago Kemppi was using matched 3xparallel IGBT's and 2-switch Forward converter topology on smaller welding machines that I'm familiar with.  160kHz switching frequency.

BUT suitable IGBT's for 650v supply voltage operation (ie 800 to 1200v rating ) are lot less common than 600v rated ones.
   

[Miyuki]

15 years ago Kemppi was using matched 3xparallel IGBT's and 2-switch Forward converter topology on smaller welding machines that I'm familiar with.  160kHz switching frequency.

BUT suitable IGBT's for 650v supply voltage operation (ie 800 to 1200v rating ) are lot less common than 600v rated ones.
 

Yes at 400V It will be easy to use Infineons H5 and they can switch even at this frequency
But 800-900V ones are very limited choice
And 1200V ones are slow

[MagicSmoker]

I would like to use ETD shape core, it have relative big area to volume ratio so I can push flux swing high (about 200mT at 200kHz) with sustainable temperature rise

You sure about that? I'm working on a transformer design right now that uses a state-of-the-art ferrite similar to TDK/EPCOS N97 and at 200mT/200kHz it wpuld have a loss ratio of 1.06kW/m³, or somewhere between 2x and 5x what is tolerable. In fact, the ideal flux swing is more like 100mT, at which point the loss ratio drops to 165W/m³ (the power of the B exponent at work), but that still is enough to cause a temperature rise of 67C over ambient (which happens to be ideal, as there is a loss minimum at 100C, similar to many other power ferrites).

Also switching is just half hard, at turn off since turn on is covered by leakage inductance at almost zero current and have just Coss losses

Eh? Turn-on is very lossy in a MOSFET-based hard-switched converter due to the energy stored in Coss. You really will need to either use SiC MOSFETs or else go with a ZVS or above-resonance topology.

what will do to switching losses driving mosfets off very hard with like -10V with minimal gate resistance and rely on diodes to clamp that spike

Makes no difference with MOSFETs. Or IGBTs, for that matter. The only reason to bring the gate negative during turn-off is to prevent shoot-through from Miller capacitance of the opposite leg turning on, but that is not a consideration in the two-switch forward.

[T3sl4co1l]

Eh? Turn-on is very lossy in a MOSFET-based hard-switched converter due to the energy stored in Coss. You really will need to either use SiC MOSFETs or else go with a ZVS or above-resonance topology.

A note, there's very little energy stored in Coss as such, but because the opposite side capacitor is in series with the supply, that voltage stacks with its Coss energy, flipping the curve: instead of delivering little power in the high-C region, it delivers maximum power there.  The super-bottom-heavy C(V) curve makes it look like a drawn-out diode recovery (it happens over 10s of volts, not ~1V), and the low side switch dissipates similar power as a result (or the switching loop inductance absorbs the energy and turns it into voltage spike, or a snubber consumes it).

Which is also a good point, with SuperJunction transistors you need to take special care of the switching loop at low current conditions.  It's not hard switching into diode recovery, but it's almost as rough.

what will do to switching losses driving mosfets off very hard with like -10V with minimal gate resistance and rely on diodes to clamp that spike

Makes no difference with MOSFETs. Or IGBTs, for that matter. The only reason to bring the gate negative during turn-off is to prevent shoot-through from Miller capacitance of the opposite leg turning on, but that is not a consideration in the two-switch forward.

Well, you can get more current through the driver+gate resistances -- this is basically why datasheet t_f figures are so much longer than t_r, there's less drop from Miller plateau voltage to driver output low voltage.  It's not usually necessary, but it can still help you get a little bit more performance.  (More often, you're burning switching speed with additional gate resistance anyway.)

Tim

[MagicSmoker]

A note, there's very little energy stored in Coss as such...

Yes, this is an underappreciated 2nd order effect - and, to be honest, one I hadn't considered here - but I mainly referred to Coss because the OP did. That said, whether Coss stores enough energy to be a problem very much depends on the type of MOSFET and, especially, the thermal resistance from junction to case/sink/ambient, etc. Those TO-220FP devices tend to have terrible Rth-jc (typically 3C/W), so even 4-5W of additional loss can be a real headache.

Well, you can get more current through the driver+gate resistances...

Technically correct, but not remotely practical - or even desirable - in the context of what the OP wants; namely, to drive (5) switches in parallel. It would take heroic efforts just to get the transition times to 50ns, much the less the <20ns that could be achieved with a single MOSFET and an appropriately sized bipolar driver in a tight loop.

[jonpaul]

Hello from Paris:



Paralleled MOSFETs are subject to parasitics inthe  VHF range, use gate damper resistors and/or beads.

Best topology for high power designs nowadays is NOT FWD, etc but soft switching, resonant, etc.

You can read numerous papers and books on these topics. This greatly lowers switching loss and can run at 500k-20 MHz. Thus caps and transformers and inductors shrink.

Just the ramblings of a retired EE, worked since 1960s in analog and power design.

Bon journee,



Jon

[Miyuki]

MagicSmoker: You are right, I was blindly using some table from TDK and it is way off at this frequencies  upper limit is much lower, more like 150mT at max duty

jonpaul: Resonant converter with constant current output and wide output voltage? For constant voltage and steady currents they are nice, but for this application it little too exotic. I dont even find much papers about this use. And have no idea how it can be controlled. 

I might try to use one big Si MOSFET like IXFH40N85X with high allowed gate voltage and drive it really hard
It have Ugss +-30V and 40V transients

So driving it with something like +20 and -10 can get reasonable small switching losses
And keep layout really tight to limit transients

[Kleinstein]

If the power is to high for a single switch, one could consider a multi phase design. It needs a 2nd / 3 rd transformer / inductor, but could reduce parasitic inductance as smaller units are used. I am not sure if there are suitable controllers.