A pulse rating on a datasheet is a current for a time, and behind it is a temperature rise: the junction, the silicon die inside the transistor, carried from one temperature to another. How hot you are willing to let it get is your decision, not a property of the part.
Which two nodes the rise is measured between is also yours to decide, and it rarely gets written down. A rise quoted without its two nodes is not yet a budget. This article explains where a pulse rating comes from, how to choose the rise for a coil's bridge, and how to check it against the transient thermal impedance curve on the datasheet.
What it is and how it works
A semiconductor die is a small thing that gets hot quickly. While it conducts it dissipates, and that heat leaves through a stack of materials: the die itself, the solder that attaches it, a copper baseplate, a layer of grease, a heatsink, the air. Every step in the stack resists the flow of heat, and every lump of material in it stores heat.

Max, in a thread on IGBT temperature, offers an analogy that makes the rest easy to follow. Temperature is voltage, thermal mass is capacitance, thermal resistance is resistance, and the dissipation in the die is a current source charging the first capacitance it finds. Each larger mass hangs off the one before it through a resistance, so the die's temperature is a node voltage part way along a chain. In his words, "the whole system is a low pass filter".
In practice this means the length of the pulse decides how much of the chain is in the circuit at all. A pulse shorter than the first time constant charges the die's own heat capacity and gets no further. The heatsink plays no part in that, and neither does the grease. Push for long enough and every capacitance is charged, the chain settles, and only the far end of it, the air, matters. A datasheet therefore carries two different things about heat: one static thermal resistance in kelvin per watt, and a curve of transient thermal impedance against pulse width.
That gives two separate questions. Whether the average power gets out is answered by the heatsink and the static resistance. Whether the peak stays under the limit is answered by the die's own heat capacity and the transient curve. A design can pass either and fail the other, and a temperature reading on the heatsink only tells you about the first.
A pulse rating comes out of the second question. Read one backwards. Somebody picked a starting condition, picked a temperature they were willing to end at, worked out the power the device dissipates at some current, and read the time off the transient curve. The rating is the result of that arithmetic. Move either temperature and the rating moves with it, though nothing about the part has changed.
What it does in a coil, and your design choices
A coil is an extreme case, because the bridge is off almost all the time. A bang lasts a few hundred microseconds and there are a couple of hundred of them a second, so the bridge conducts for something like six per cent of the time [derived: 300 µs at 200 bangs a second]. The ontime is the typical figure from the interrupter is the note, "typically no more than about 300 µs a note". The rate is only an illustration, as that article says itself, "There is no best rate, since the rate is only the pitch", and its own worked duty is 5 per cent rather than six. The average power is small and the heatsink copes easily, but inside those few hundred microseconds the die heats up in a way the heatsink never sees.
Jimmy Hynes summed up the design problem in two sentences on the Tesla list while working out his second DRSSTC (pupman). The problem was letting the die absorb more energy during the short bursts, and getting the heat out of the heatsink was not a problem at all. He proposed putting a piece of copper on top of the die for it, and said in the same breath that it did not have to move any net heat out, only add thermal capacitance. This is where a coil's thermal design differs from a converter's: you are sizing something to hold the heat, not something to remove it.
Which two nodes your rise spans
Before deciding how large a rise to allow, say what it is a rise of. The one published figure for this shows why.
davekni gives it as part of an argument. He says that for most DRSSTC designs the transient rise is larger than the heatsink's rise, and the figures illustrate that claim: "If pulse rise is say 100C, and heatsink is at 50C, then peak junction temperature is 150C." That rise runs from junction to heatsink, so it takes in the package and the mounting joint as well as the die. The word "say" is his.
It is an example, not a budget he recommends, and the useful lesson is that the rise is a choice. The same 100 read as junction to case, or junction to ambient, is a different budget, because a different share of the chain sits inside it. What a large swing does to a device's life, the swing rather than the peak, belongs to the bridge article.
Scaling a pulse rating to a shorter pulse
Two builders worked a pulse rating in public on the Tesla list on 1 October 2004, on two different bricks, and a third answered with a model instead of a number.
Steve Conner took the datasheet line at face value and pushed it as far as it would go (pupman). For a Semikron SKM300GB123D the maximum pulse rating is 600 A for 1 ms, and the datasheet says the pulse width is limited by junction temperature. From that he proposed the temperatures behind the rating, and labelled the proposal as one: "That suggests that 600A for 1ms will raise the junction temperature from 25'C to 125'C", with the rating falling to 440 A at an 80 °C case being what "reinforces this hypothesis". Then he treated average current times time as the invariant, so "1800A average for 333us is the same as 600A for 1ms", and carried that through a full bridge to an energy per bang.
He flagged his own assumptions in the same message, so read the message itself rather than lifting the numbers. The arithmetic assumes the forward voltage does not change with current, and in practice "the temperature may be higher with shorter pulses as the heat has less time to escape to the baseplate."
Jim Lux answered with the model rather than the rating (pupman). He asked whether the manufacturer publishes a two or three node model, die mass, theta jc, case mass, theta cs (the thetas being the thermal resistances from junction to case and from case to sink), because with those masses and resistances "it's just like a cascaded RC network, and you can calculate the time constants." That is the RC chain described above, offered in place of the datasheet's front page. He produced no number of his own.
Jimmy Hynes had already taken that route, on a different part (pupman): "I didn't use that number for my calculations, I looked at the transient thermal impedance charts on the datasheet. Those make it look even better than your method." His device was a Mitsubishi CM300DY-24H and Conner's was the Semikron, so the charts coming out kinder is one part read against another part's rule, not one datasheet read two ways. The exchange never resolved it, and could not have. The answer is the transient impedance curve for the part in your own bridge, at the pulse width you run.
Here is our own reading of why a rule like Conner's is a rule and not a law. It follows from a sentence on IGBT or MOSFET. Current times time is the adiabatic answer, the one with no heat leaving the die during the pulse, for a device whose loss is V_ce·I, because then the energy is proportional to I·t. A MOSFET's loss is I²·R, so its adiabatic invariant is I²·t instead. Halve the pulse and an IGBT takes twice the current for the same rise, a MOSFET only √2 times [derived from those two loss laws]. Neither rule holds past the first time constant in the chain, where the die starts genuinely sinking heat into the copper. Conner's own caveat says the same in other words.
What it takes to build the thermal model
Hynes built his in SPICE and said what he had to find (pupman). All the values for the IGBT came from the datasheet except the thermal capacitance of the baseplate, which he calculated, and the heatsink's, which he got by weighing it. The output stage of the simulation was a gain block converting joules to degrees, which is the same idea drawn as a schematic.
He also planned to measure it live. The plan is useful, though the message is in the future tense throughout and nothing in it had been built yet. He rejected a thermistor on the die on paper: its thermal time constant was too long, and the differentiator needed to compensate would amplify the noise too much. He proposed an infrared photodiode instead, which measures emitted infrared with "only 1uS of delay", so a bang could be cut short from inside itself. He also noted the mismatch, that the wavelength of those diodes is "near 900nm, which is a little short for the temperature I'm looking for (~100C or so)".
Measuring die temperature, and what the reading means
davekni makes the same split between average and peak in the same thread. For average power, the top of a TO-247 case centred on the die is a fair approximation, and he gives one reading of his own for it. For transient thermals during an enable pulse, "most people rely on calculations", and the two measurement options he offers are not equivalent. The first characterises a temperature dependent parameter of your own part and uses it as a transfer function. Only that one carries the warning given further down: what it returns is the die's average temperature, while the specification is written for its hottest point. The second is half measurement and half calculation, and it ends back at the datasheet, scaling the measured average power by one over the duty cycle and reading the rise off the transient thermal resistance graph.
He also puts a condition on the calculated route that is easy to miss. The estimate presumes the case to heatsink resistance, the grease layer in the stack, is low. If that joint is poor, the rise you calculated sits on a baseline you have not measured.
The numbers
- The rise is the free variable, and it does not count the same for every device. The thermal impedance you are allowed to spend is that rise divided by the power, so it moves in proportion to the rise. The current you get for it does not. At
V_ce·Ithe current moves nearly with the rise, and atI²·Rit moves as the square root [derived from the two loss laws above]. Whatever rise you pick gets an IGBT builder more current than a MOSFET builder. - The calculator and the ΔT field. Switch thermals opens on the site's reference coil: a MOSFET full bridge of eight devices, two per position, 160 A peak, a hot
R_DS(on)of 0.19 Ω, 2 ms ramps 70 times a second, which is 14 per cent duty. Set the current to steady, the worst case, and that is 40 A RMS in each device, 304 W during the burst and 42.6 W averaged [derived from those inputs: 160/2/2, thenI²R, then times 0.14]. At the 100 K the ΔT field arrives set to, the transient budget is 0.329 K/W and the page calls the requirement moderate. Put 30 K in, change nothing else, and it is 0.099 K/W, and the same page tells you to go and read the datasheet curve because the staticR_th(j-c)will not do [both derived from the page's own arithmetic]. Same bridge, same current, and one field changed the verdict. Two caveats before you take this example anywhere else. Its loss model isI²R, a MOSFET's, while the rise discussed above came from an IGBT thread (the page handles an IGBT too, by itsV_CE(sat)). And a 2 ms burst at 14 per cent duty is a ramped coil, not the few hundred microseconds this section opened with. - Steve Conner's chain, and where he stopped. 600 A for 1 ms, 440 A at an 80 °C case, a proposed 25 to 125 °C, "1800A average for 333us", and from a 650 V bridge a peak of 5902 A, an RMS of 4173 A and 1149 J per bang, which he labelled an absolute theoretical maximum in capitals. He then declined to use it. The short circuit rating is 2000 A and not guaranteed past a thousand pulses, so "I would keep my peak current down to 1000-1500A and that limits me to 50-70 joules." The derating he states on current is a factor of 3.9 to 5.9 [derived: 5902 over 1500 and over 1000]. His joule figure is not for that same bang. His own energy expression at 333 µs and those peaks gives 195 to 292 J [derived: 650·(4/π)·(I_pk·0.707)·333e-6, which at his 5902 A returns 1150 J against the 1149 J he printed], so the shorter bang behind 50 to 70 J is not stated anywhere in the message. Take the current limit from him, not the energy.
- The case is not the die, and one reading is not a calibration. davekni measured "~140C on the top of the case when die temperature was 150C", on a TO-247 of his own, at average power, which is the only case where he says the method works at all. That is 10 °C of offset on one part [derived: 150 minus 140]. It is the difference of two approximate numbers and it grows with dissipation, so what transfers to your bench is the sign and not the size. A case reading is a floor under the die.
- How far past its rated peak a builder will run a part. Steve Ward, on the Tesla list (pupman): "If only the bigger IGBTs could withstand such abuse. I only run my CM300s at 2X their peak rating." The abuse he means is in the message he was answering (pupman), where Terry Fritz reports an IRG4PH50UD, a "24 amp rated TO247AC IGBT with diode", "happily eating 700+ amps" against the "180 Amp peak rating" it gets because "it uses a great big die in a small package". Those two multiples are nothing like each other. The reason Fritz gives for his is the die rather than the package, and the die is also where a short bang's heat goes. Note also where Ward's routine sits. If his peak rating is the datasheet's pulsed figure, then twice it is where switch thermals puts davekni's rule of thumb for where devices fail.
What to watch out for
What goes wrong
- Devices die weeks in, well below
Tjmax, with nothing wrong on the scope. The cause is the size of the temperature swing, not the peak. The swing stresses the attachment between die and heat spreader, and the wire bonding with it. Weston and Hydron both say so in the same thread, and Hydron's version is the more useful one: the swing becomes problematically large long before you reachTjmax, and that drove a lot of his QCW design. The mechanism is covered with the bridge. - The heatsink is cold and the bridge fails anyway. You sized the static path. At a few hundred microseconds the heatsink is not in the circuit, and what killed the device was the first rung of the chain.
- A device that was fine on the bench dies on the arc. The pulse rise did not change. The baseline under it did, because the arc run has a higher duty and the heatsink came up under it.
- Same current, longer ontime, sudden failures. Transient impedance rises with pulse width, so the same power for longer is a larger rise, and the ontime is the term you moved. Raising the rate gets you to the same place the other way, through the average, and the interrupter page has the duty arithmetic for both.
- You measured 140 °C on the case and called it safe. In the one reading davekni reports, his die was at 150 °C while his case read that. The offset is his part's rather than yours, but the direction holds for any part.
- The budget passes on paper and the device still dies. Check whether the switching loss field was left at zero. With soft switching working it is close to nothing, and when it stops working there is no predicting it, so a conduction-only figure is a floor rather than an answer.
- A paralleled pair with one device much hotter than the other. That is gate drive mismatch, not thermals, and IGBT or MOSFET covers it.
Where next
- Switch thermals does the arithmetic above. Set the current to steady, then change the ΔT field and watch the verdict move, because that field is the main point here.
- The bridge and what kills it, for the swing as a failure mechanism rather than as a budget, and for when the answer is that the pulse does not fit in one die at all.
- IGBT or MOSFET, for the loss law that decides how a pulse rating scales.
- The interrupter is the note, because rate and ontime are the two knobs that set the average.
Every figure above is quoted from a source or marked as derived, with its inputs. We measured nothing here ourselves, and none of it replaces the transient impedance curve for your own part.