-
I've been wondering for YEARS now if this is actually a thing or what.
Turns out it is, and I finally have the keywords to search on it! And it does have consequences in-circuit! Not just a measurement discrepancy.
The smoking gun: I measured this awkwardly drawn plot back in 2013,
A later revisit (2017) produced these data,
STP19NM50N was a SuperJunction type (ST's MDmesh II) power N-MOSFET. Like most SJ types (pretty much anything >= 200V since mid-2000s to 2010s), it shows a suspiciously abrupt drop in Coss at low voltages (typically at 10-50V for a 600V-class part). But I was never able to reproduce that curve. All my measurements (based on dV/dt) showed the fatter curve.
So what's the deal? They're lying, right?
This reveals what I was missing:
https://www.slideshare.net/MichaelHarrison96/coss-hysteresis-in-advanced-superjunction-mosfets-apec-2016-presentation-compressed
particularly page 7 fits everything absolutely perfectly.
If I understand this correctly, it's effectively a recovery effect; not recombination mediated, as the minority carriers in a diode/BJT, but a majority carrier effect, migrating through the region as it depletes. Effectively the junction capacitance has dielectric loss, or high ESR, but only transiently as it's discharging from the very-high-C state (where the depletion region zig-zags up and down the SJ pillars) to the low-C state (fully depleted, high-Q silicon insulating between D/S).
https://superlab.stanford.edu/files2download/TED2018_paper_Grayson.pdf
This explains it as an RC equivalent circuit, noting that the pillars effectively charge in series so making an RC (lossy) transmission line like structure. And the C's are voltage-dependent, of course. More measurements are also provided, showing the frequency dependency.
...Gosh, I wonder if this applies to those "schottky" diodes, the kind in higher voltage ratings that have a t_rr spec (that is definitely there (I've measured that too), and is independent of temperature). I bet it does. I'm not sure what all technologies are applied in those things, but they must be using SJ in at least a few families. The same effect will look exactly like recovery loss; though, the charge should be dependent only on ΔV and dV/dt, not If. But it does depend on If; mhh, perhaps some conductivity modulation and recombination taking place then? Oh, maybe my measurement jig produces correlated dV/dt and If, that could be too. Ooh, I wonder what dynamic breakdown looks like on those; can they avalanche lower during t_rr (as regular PN diodes can)? Something else to try...
Tim -
All datasheets have at least one dodgy statement or sneaky chart.
-
Data sheets are full of lies, old wives tales and historic quotes to please ancient engineers. Think of them as toilet paper for CYA purposes.
Like pretty much every CMOS device before about 10 years ago (and most after) quoting only TTL environment data. Seriously, who is feeding old 1970's TTL devices from 21st century high speed CMOS memory? Keep yanking, I may cum around...
Or how about the specification for testing power supplies. Scope limited to 20MHz BW? Gim'me a freakin' break! How is that pertinent to today's GHz world? "Oh well sonny, that's the way it was first done in 1966!"
Yeah, take most specification and data sheets with a grain (or two) of salt and a bottle of Scotch. -
I was about to start a topic like that. I'm in the market of small signal fets, and I'm looking for those devices with lowest gate capacitance so I can easily drive them by an opamp. The thing is, according to my transistor tester, gate capacitance often doesn't match typical value. Like FDV301N measured 200pf instead of 50pf "typical". The DS provides only typical value. I had same thing with other devices. Like, one device measured 80pf instead of 25 or 30.
Idk how much I can trust transistor tester on that, but it seems for some devices it does display what's close to typical value. -
It is a little suspicious that they don't show a test circuit for measuring the COSS, but I would assume that they did it with zero gate bias, since that is not a variable. But then I wouldn't expect COSS to vary with current either, tricky. Thanks for the info.
-
Or how about the specification for testing power supplies. Scope limited to 20MHz BW? Gim'me a freakin' break! How is that pertinent to today's GHz world? "Oh well sonny, that's the way it was first done in 1966!"
Higher frequency noise is largely irrelevant because it is easily passively filtered. A 10 or even 5 MHz measurement bandwidth would be better, but 20 MHz became the standard. Some old oscilloscopes do use 10 or 5 MHz instead of 20 MHz, and then a correction factor has to be used for fair comparisons.
-
I was about to start a topic like that. I'm in the market of small signal fets, and I'm looking for those devices with lowest gate capacitance so I can easily drive them by an opamp. The thing is, according to my transistor tester, gate capacitance often doesn't match typical value. Like FDV301N measured 200pf instead of 50pf "typical". The DS provides only typical value. I had same thing with other devices. Like, one device measured 80pf instead of 25 or 30.
Idk how much I can trust transistor tester on that, but it seems for some devices it does display what's close to typical value.
That one is fine, but take note of all the conditions. For VDMOS, Ciss depends on Vds. In fact, in the Vgs < Vgs(th) region, you get a delightfully variable capacitor, albeit not a high-quality one (depending on RG; usually a few ohms though), and a pretty fair voltage range (20Vpp or so) with high linearity.
More specifically, Cgs is ~constant and Cdg varies. You can see this in the plots, Ciss = Crss + a constant, more or less. It's normally a small difference a log plot, but it's there. Downside, the large fixed Cgs rather limits tuning range, of course.
This is probably surprising if your state of knowledge started with lateral (Ciss depends on Vgs, Vgd), or you're inferring from a similar point (in general, MOS capacitors are voltage-dependent, why not MOSFETs?).
It's also funny to me, because a lot of SPICE models even modeled Cdg as dependent on Vgd, even to the point of making a virtual rectifier and dependent capacitor circuit (i.e. on the assumption that C(Vdg) = C(-Vgd)), but nah it's a three-terminal device with Vds biasing it, not Vgd at all. Maybe that was itself a better approximation, or even a whole truth, in older types, which the models were invented for, but then applied to HEXFETs and then trench MOS where it isn't very realistic at all. No idea.
As for the claimed effect, it might not affect small transistors, depending on what design they are; AFAIK, SJ hasn't come to very low voltages yet (it requires a high aspect ratio to work, so if you're making the pillars short (thin depletion region, low voltage rating), they have to be incredibly narrow, but can't be made finer than the tech node feature size), and those are still more or less stock VDMOS, though there are still many incremental developments being made that I don't know about. And old designs that are never updated, aren't, so there's always IRF540 etc. if you need something slow, cheap, modest power, and only mostly nonlinear.
For something like a linear circuit, the stability effect of Crss and Coss variation tends to be stronger, forcing some minimum compensation setting; in contrast, the high gm tends to reduce the impact of Ciss, i.e. perhaps compensation dominates. And consider an R+C at the output to ballast load reactance, if this is something general like an electric load.
Tim -
Higher frequency noise is largely irrelevant because it is easily passively filtered.
Yeah.
No, sorry, not buying it. If it was "easily passively filtered" and they were serious about producing an acceptable product they'd include that easy-peasy filtering in the design and not hold the world to such preposterous limitations. Complete fail.
Yeah, time to move into century 21 and offer products (and specifications) that work "out of the box" in today's applications. If I buy power supplies from a power supply company for my product, I should not have to hire a power supply engineer just to clean up their sh!t.
Oh, and if anyone is still using 10 or 5 MHz scopes today to do QA, WTF? Are they somewhere deep in the Amazon? What products are they making? Mashed potatoes?
Sorry, no excuse. -
Is this related to Vds sweep rate? I don't think they state that, usually, but at super low or super high frequencies you're going to see different behaviors owing to the transistor not being a point object.
Sort of. The combination of slow sweep, small amplitude test at modest frequency (but 1MHz is in the ballpark where the effect is strong), and most of all measuring just the reactance (ignoring resistance) and converting that to capacitance, gives the plot shown. As far as point-iness, the fact that it's a die of some ~mm across isn't significant (by design, of course!), but it seems the SJ pillars leave some non-point-iness in the vertical cross-section of the junction. Well, inhomogeneity moreso than points or the lack thereof, but I suppose that's still something to point to.
Mmm, Nonpoint, what an interesting Development this is.
Tim -
at some point I agree with the bandwidth limitations on measurement because if a company had to provide all that shit we would never get any product because it was going on the RF design beuro. Alot of customers DGAF about anything more then HF. You start dealing with serious baloney because of wiring anyway, it will never work for anything but the most simple implementation, as soon as you do some systems engineering it won't be relevant, and it will just give people a false sense of confidence. And keeping old equipment in service and relevant keeps costs down and allows easy entry into the field, which is wonderful for price and availability. I like options.
I love having lots of power supply options, in different form factors, then at least you have a chance to make your stuff work without mechanical redesign, as soon as you are doing something tricky and non standard. Ah so someone slapped together some power supply thats doing something a little weird, but it looks like I have room under system element #14 for maybe fix it, and its cheap, wonderful news, I get to do some problem solving, and rewarded for exploring technology
and if that stuff was specified, how accurate would those specs be compared to the other specs in regards to drift, aging, damage, etc? might be seriously dodgy. Some people modify their subassemblies too....... too much bounding is no good. I just accept that fets are tricky and they have their rewards if you have the patience to study them -
Oh, and if anyone is still using 10 or 5 MHz scopes today to do QA, WTF? Are they somewhere deep in the Amazon? What products are they making? Mashed potatoes?
No, these oscilloscopes have 5, 10, or 20 MHz bandwidth limits, but are much faster otherwise. The bandwidth limit is associated with the highest vertical sensitivity combined with the noise of the input stage. So for instance the Tektronix 2225 is a 50 MHz oscilloscope with a 10 MHz bandwidth limit because its input sensitivity is 500 microvolts/division. The 100 MHz Tektronix 7A13 has a bandwidth limit of 5 MHz because its bootstrapped input design has several times higher input noise. The 1 MHz Tektronix 7A22 has bandwidth limits down to kHz because its input sensitivity is 10 microvolts per division.
20 MHz is about right for 1 or 2 millivolt per division sensitivity and good input noise.
-
Ah, okay. It was not fully clear in your first post.
But I stand by my claim.
If I were to make a PS for myself today (and I sometimes still do) I make it as clean as I can. If I were to make them commercially, I'd do the same. Noise filtering would be in place.
For instance, I recently have undertaken a project to improve a low-end DDS AWG and one of the targets for that improvement is the PS. I don't want to add a PS and another board cleaning up the mess the PS makes. I've tested 3 SMPS units. Only one passes muster. The Mornsun. It beats it's specifications even when tested with a 500MHz scope without any BW limits enabled. Guess which one I will use? Also guess where I'll go in the future to get my SMPS modules?
Mornsun publish their specifications citing that ancient 20MHz limit (because old engineers, those in charge, want to see that. It's what they are used to), but they obviously design for and test for much better performance. Meanwell (nice name!
) are also pretty good and handily beat their specifications too. Many only meet them with a crippled scope looking at them. No, no .. ah me no buy! (as said by Consuela - Family Guy)
-
are you sure that efficiency is not more important on a large scale? everyone is trying to think green now, not quiet.
Replace all those 82% switchers with 92% and suddenly you are saving ALOT of energy on all the useless dodads doing nothing at all, all the time. Digikey goes up to 98% but there is hardly anything there.
I think its very important, look at for instance dave jones video about smoke alarm power consumption! typically not that kind of power converter, but still... adds up quickly. -
Mornsun publish their specifications citing that ancient 20MHz limit (because old engineers, those in charge, want to see that. It's what they are used to), but they obviously design for and test for much better performance. Meanwell (nice name!
) are also pretty good and handily beat their specifications too. Many only meet them with a crippled scope looking at them. No, no .. ah me no buy! (as said by Consuela - Family Guy)
With bandwidth wider than 20 MHz, broadband white noise increasingly conceals the lower frequency noise which is more significant, so improving the noise that matters has less of effect on the measurement.
An example of this is operational amplifier low frequency noise, which is specified separately from broadband noise. If you make a 20+ MHz measurement, then the low frequency noise is completely concealed, but it is the noise from 0.1 to 10 Hz which is important in precision DC applications because it is so difficult to filter, while higher frequency noise can be ignored.
A spectrum analysis returning noise density could be much better, but was too difficult to do in the past and not necessary.
-
MOSFET Coss related losses are something I recently learned about as well. One of the best public papers I found so far is this one https://www.aramis.admin.ch/Default?DocumentID=69538&Load=true which quantifies and compares the effect for different technologies and device families, both for the frequency-independent (hysteresis) and frequency-dependent (Coss delta tan) components of the loss.
-
This reveals what I was missing:
https://www.slideshare.net/MichaelHarrison96/coss-hysteresis-in-advanced-superjunction-mosfets-apec-2016-presentation-compressed
particularly page 7 fits everything absolutely perfectly.
There is a video of this presentation on youtube, definitely worth checking out.
-
My trust in datasheets in general has reduced over the years. I still "trust" the datasheets from big names (Samsung, TI, Analog Devices etc), but I have a rule to only trust the datasheet values of parameters that I don't care about. For everything that I do care about I want to get the part on my bench and test it myself.
Just a few months ago I was designing and making a battery powered temperature logger for personal use. I bought a Chinese 3V3 LDO specifically because of its low quiescent current (1uA advertised). I put the part into my design and the quiescent current actually was 1mA.
Needless to say, runtime projection dropped from 7 years to 18 days because of 1 part. -
My trust in datasheets in general has reduced over the years. I still "trust" the datasheets from big names (Samsung, TI, Analog Devices etc), but I have a rule to only trust the datasheet values of parameters that I don't care about.
I have not trusted TI datasheets for decades. Does anybody remember when they released new integrating ADCs with the feature of not returning positive and negative zero, which still returned positive and negative zero, and then refused to correct their datasheets and advertising? What about confusing common mode rejection versus frequency with power supply rejection versus frequency? And for a long time their noise specifications were rubbish, and maybe they still are.
I just bought some TI chopper stabilized amplifiers for testing because the datasheet specifications for noise seem too good to be true.
-
I have not trusted TI datasheets for decades.
I came to similar conclusion: they try very hard to make datasheets look good. The devil is in details. Like, they try to make their cmos opamps attractive and comparable with bjt opamps in terms of noise and and offset.
I still prefer TI over other vendors as their datasheets look comprehensive and provide lots of plots and info. The other vendor I like is LT/AD, but they are way too expensive, imo. -
I had an issue some time ago using an SG3525 PWM controller. The datasheet said the discharge resistor value for the sawtooth RC oscillator could be all the way down to zero ohms. Trouble is, if you do that, then if the error amplifier is maxed out the sawtooth waveform goes slightly higher and lower ever second cycle, leading to unequal output pulse pulse widths with ensuing problems.
-
All datasheets lie- dont take anything at face value!