Verifying the Datasheet and SPICE Model

A few more measurements this evening.

ST's MDmesh II SPICE models are terrible, and their capacitance plots are characteristically wrong:

formatting link

I had previously measured this,

formatting link
which shows the capacitance tails off not quite so aggressively as the datasheet suggests. But it ends at 90V, which is inconvenient.

I've improved the measurement, using a drain charging resistor (100k), extracting C(V) from the waveform. This covers the whole (500V) range.

formatting link
Cursor shows the 80% crossing mark, which for 100kohms, you take time (us) *

6.21 = Coss (time equivalent) (pF). Their figure is 202pF, I get 131pF.

Coss:

formatting link
including best-fit diode+cap model parameters.

They don't have a SPICE model for this exact part, but the nearby STP21NM50N (which is a somewhat bigger and less well optimized design, now obsolete) gives a figure of 270pF (time equivalent); SPICE says 389pF. It's wrong by the same amount in the opposite direction. :-)

Using the same extraction on the SPICE model (or simply inspecting the code) yields this:

formatting link
Well, it's correct at 21V. Clocks twice a day and all that.

The parameters used for the fit (red) are, CJO 887.8943 pF m 0.267762 VJ 0.942384 V Clin 124.6118 pF which are in the ballpark of the what appears in the code. I can conclude the extraction method works fine.

I like how the SPICE models proudly proclaim "PARAMETER MODELS EXTRACTED FROM MEASURED DATA"...

How about something they actually got right?

formatting link
Probes: Ch1 100x, Ch3 10x. Ch2 shows gate timing. When on, 1.045mA CCS charges gate; when low, a switch shorts it to GND. 13.3A drain load.

(The drain supply was fun: my benchtop HV supply is a pissy little thing, topping out at about 10mA. Function generator wired as pulse generator afforded the low duty cycle, which turned out to be 30us on, ~100ms off. For you analog afficionados, this measurement would be needlessly difficult without a DSO. Not to mention the trace averaging. Nice!)

Anyway, I measured 32nC total Qg, datasheet says 34. Not bad.

The '21 SPICE model measures 68.3nC, datasheet says 65. Also not bad.

Transfer function: extracted from same waveform. Surprisingly close:

formatting link
Of course mine stops at 13A because Vds isn't actually constant.

SPICE model not so hot:

formatting link
But, looking at this closely... the datasheet does actually specify g_fs =

12 S. The model's slope is spot on 12 S. The datasheet curve shows only about 8.8 S! Who is right? :-)

I wonder if they stopped tabulating g_fs because they had consistency problems measuring it, or something. Their own plots contradict the 12 S figure...

Tim

--
Seven Transistor Labs, LLC 
Electrical Engineering Consultation and Contract Design 
 Click to see the full signature
Reply to
Tim Williams
Loading thread data ...

Tim, the super-junction MOSFETs I've measured have in fact shown abrupt drop-to-almost-zero plots, just like those shown in the datasheets. I make measurements with an HP 4280A precision semiconductor capacitance meter, in a fixture modified to allow for Vds voltages to 500 volts. Measurements I've made with the 4280A match the results I've gotten on my other RLC instruments.

I've placed a few of the types of MOSFETs you've been playing with on my next Digi-Kay and Mouser orders, and will try to get measurements on them.

--
 Thanks, 
    - Win
Reply to
Winfield Hill

My C measurements are pulsed (full edge waveforms), so it may very well be the case that there's some strange difference between that and the RF small signal version (which I presume is what your tools perform).

But if that were the case, it wouldn't be very useful, would it? -- you're only exploring these regions of operation in switching applications (to the extent that, that's what the parts are intended for), so the edge method /has/ to be the more correct option, if there be a difference at all.

I'd had some Infineon parts as well (of the same "abrupt-near-zero then slight rebound" performance), but, checking, it seems I've used them all up. Maybe I can pull one out of something and test.

Mind also that, what I'd measured still doesn't quite map to what models the right waveforms. So there's three sides of it, even.

Cheers!

Would you like a copy of my data and/or models to compare as well?

Tim

--
Seven Transistor Labs, LLC 
Electrical Engineering Consultation and Contract Design 
 Click to see the full signature
Reply to
Tim Williams

The measurements are at 1MHz, 20mV and they're totally legitimate, that's why that's the measurement standard. They show things you have trouble seeing, for example when Coss is very low the trace is too fast and it's hard to use dV/dt to make an accurate measurement, and when Coss is is very high, below 5 to 15 volts, dV/dt slows and it's hard to see what's happening. I notice in your scope trace comparisons a considerable deviation below 25V. An effort to optimize for one condition will make it even harder to see the other.

A serious issue in high-frequency high-power MOSFET switching, half-bridge, etc., is how long it takes the MOSFET to fully turn on, lowering I^2 R losses, and even more important, how long it takes to fully turn off, lowering switch shoot-through losses. Fully analyzing this scene requires accurate MOSFET capacitance measurements and modeling. I find the former easy, the latter is a mess. For example, gate spreading resistance issues mean that remote portions of the FET's gate take longer to discharge.

Some MOSFETs are really bad in this respect; their datasheets are of no help, it's a bench-test matter. This can be modeled with a second smaller parallel FET, with a slower gate. The big one turns off quickly, the small one hangs on a bit at turnoff. One can also see this by observing losses in excess of Coss V^2 and integrated I*Vds during switching. One approach is to just toss out any offending types, and continue the search for a better part.

Yes, that'd be a good idea. I'd like to explore how you run your spreadsheet analysis. Maybe it can be beefed up to better show the regions that I think are important.

--
 Thanks, 
    - Win
Reply to
Winfield Hill

I didn't show because it's boring :)

formatting link

And, indeed, that's the reason I took this waveform at low voltage: to account for error in the high voltage measurement. I think my 100x probe compensation is a bit batty too.

Yeah, modeling diffusion across transistors would be ridiculous.

Example: I made a 2N7002 cascode amp, just for kicks. It's -3dB at 50MHz or so, but continues to have useful gain up to 200MHz or more. It's not a single pole response!

Physics suggests splitting Cgs and Cdg into pieces, distributing them along a chain of Rg's. And probably distributing M's along the taps, as well. (Just as you said (snipped), for the case N=2.)

Such a model would simulate slower, too.

Speaking of 2N7000/2, the models are quite inconsistent on this matter. Some use a single big resistor, around 300 ohms, which is simply untrue. Others use a teensy resistor, which is probably more realistic in terms of gate input impedance, but ludicrous on speed.

You could capture the small signal gain by using a diffusion network (RCRC), single MOS, and maybe doing a compromise with Cdg. But yeah, that'll miss switching loss and stuff.

Alright. I'll send them off.

Thanks, Tim

--
Seven Transistor Labs, LLC 
Electrical Engineering Consultation and Contract Design 
 Click to see the full signature
Reply to
Tim Williams

I wouldn't call it diffusion, exactly. And thinking of small dies like a 2N7002 isn't too helpful either.

Take a look at Fig 3.105, page 201 of AoE III. Here we see a power MOSFET from a different view, a vast landscape spread out over a huge region, with puny gate-connection channels squeezed in here and there. Realizing there might be distant regions less-well served is obvious. This particular Motorola photo is of an old design. Newer MOSFET designs, with greatly narrowed line widths, may be at more risk.

--
 Thanks, 
    - Win
Reply to
Winfield Hill

Some manufacturers are paying extra attention to gate resistance, and presumable the diffusion issue (as Tim called it) as well. For example, Infineon's CoolMOS IPP60R520 had an original CP version with Rg = 1.3 ohms, but the new E6 and C6 versions have Rg = 7 and 13 ohms. All other specs are identical. Superjunction MOSFETs can suffer from "too fast" switching and high Rg can provide dV/dt dI/dt control and reduce ringing. I'm not sure if their approach is superior to simply adding an external gate resistor, but they subsequently discontinued the CP version.

--
 Thanks, 
    - Win
Reply to
Winfield Hill

Those things can /scream/. Pushing a ~6N60 as fast as a UC3842 can go, it still finds plenty of time (some 50ns) to squeal at 300MHz. Ferrite bead on gate and source fixed it.

One of these days I'll set one up as an RF amp and see how that goes.

Tim

--
Seven Transistor Labs, LLC 
Electrical Engineering Consultation and Contract Design 
 Click to see the full signature
Reply to
Tim Williams

Different 2N7002s from different manufacturers behave differently. We make some 50-volt 50-ohm pulsers that only work right if we use one particular mfr's part. It switches a lot harder than the data sheet suggests.

formatting link

Incidentally, a Boonton 72

formatting link

is a great C-V meter. It has bias inputs on the back. They are available from brokers or ebay.

--
John Larkin         Highland Technology, Inc 

lunatic fringe electronics
 Click to see the full signature
Reply to
John Larkin

Though to be fair, 2N7002 is pretty honkin', as "small signal" types go. It has ~3x the capacitance of a 2N4401, and can sink slightly more current. Given the performance difference between BJTs and MOSFETs, that means its size should at least be close.

It's not clear how much diffusion (spreading, etc.) applies to that, or other types, but on the upside: if they go to the effort of building a fractal interconnect, that gives equal resistance to all points, and therefore eliminates diffusion -- it's a simple pole. We only need effort if they're being lazy (like below), or for the last few micrometers within each stripe.

Hmm, looks like stripes, joined with a grid of metallization. (The stripes could be poly or metal, I don't know.) It's not clear what type of design it is: lateral, VMOS, or whatever.

You have the equivalent circuit where distant stripes are fed by a longer distance on the grid, but also where stripes themselves have length. The fastest elements would be the ends of strips nearest the pad, and the slowest would be the middles of strips in the corners.

A complete model would have to look at each zone, but an approximate model can probably do well with a simplified equivalent.

What might be interesting is, a precision plot of gate input impedance, versus frequency, for various transistor designs. You'd be able to see things like pitch and thickness of the interconnect grid, stripes, and if they're structured / geometric / fractal or not.

Hmm, for the precision required to identify most of those things, the length, size, number, and even exact orientation, of the wire bond(s) might matter...

Tim

--
Seven Transistor Labs, LLC 
Electrical Engineering Consultation and Contract Design 
 Click to see the full signature
Reply to
Tim Williams

Den mandag den 2. oktober 2017 kl. 17.44.22 UTC+2 skrev Tim Williams:

formatting link

Reply to
Lasse Langwadt Christensen

ElectronDepot website is not affiliated with any of the manufacturers or service providers discussed here. All logos and trade names are the property of their respective owners.