LM3409HV: Expected efficiency.

But is my expectation of "on the order of 90% efficiency" a reasonable expectation ?

Is there something that glares out at you with  the choice of  components. Any "oh, that's obviously" the wrong 

I carefully studied the document you pointed at.  Taking into acount the "givens" 50V Vin, 330kHz switching, 0.6V across the diode, an on-percentage of about 60% (30/50),  0.2A of current, 0.3Ohms RDSON and 1 nC of gate charge until the miller plateau (0.8 or less from the datasheet), 1.5 of miller charge (a little less from the datasheet), 1.0A gate drive strenght (miller is at 3V, so the LM3409 has 3V to work with, so 3V/2 Ohms = 1.5A which is larger than the 1A claimed curretn, so we take 1.0A). 

I get conduction losses i^2*R * DCon = 7.2mW. (in the mosfet). 

Diode losses I*Vf*DCoff = 48mW (not in the mosfet)

Then switching on losses up to the miller plateau: t1 * I/2 * Vds= 1.65mW (in the mosfet)

The switching losses in the miller plateau t2 * .VDS/2 * I = 2.5 mW (in the mosfet). 
and lastly the gate drive losses. On the one hand that's about 6nC * 6V *Fswitch,  But if you add in the losses in the linear regulator in the LM3409 that simply becomes 6nC * Vin * Fswitch. This comes to 99mW, but not in the mosfet. 

So.... I can explain about 150mW of losses outside of the mosfet and 10mW of losses in the mosfet. I'm seeing a total of on the order of 10 times more losses. What am I doing wrong?  

I think that with simply assuming the "max current' and ignoring the ripple I've calculated a "max" or "upper limit" as opposed to the most accurate possible. But if my simpler calculations for the upper limit are still a factor of 10 away from the observed losses, I don't see the point in trying to improve the accuracy. (reminder: I calculate an upper limit of 250mW and I'm seeing 2W of losses). 

I understand that for example "gate charge losses" are incurred only once per "buck converter" and that this will spread out over more actual power delivered if the total power is larger.  On the other hand, my "small-ish" fet has lower Q-numbers than a bigger FET. But the "gate charge losses" happen in the LM3409 not in the mosfet. Once those losses are dominant, I can evaluate if it makes sense to make the VCC supply of the LM3409 with a switching powersupply as opposed to the LDO in the LM3409. (Yes, that's not trivial as the flow of current will be the wrong direction). Anyway. It is bedtime here. I'll study that document in the morning. 

I don't think so. 

In An Accurate Approach for Calculating the Efficiency of a Synchronous Buck Converer
section 2.1.1 this is explained. 

First, lets calculate the t1, t2 and t3 times times. The datasheet for my mosfet doesn't have separate Qg values for t1 and t2. This means that by using the value mentioned instead of only the value for Q(t2) I will be overestimating the power loss. 

Niow the time t1+t2 can be calculated as Q(t1+t2) / I  = Qgs / Idrive = 0.83nC/1A = 0.83ns. 

So then we calculate Esw2 = 0.5 * I * VDS * t = 0.5 * 0.2A * 50V * .83ns = 4.2nJ. Multiply by 2f to get the on+off poiwer and we get 1.4mW.  The other calculation goes similar and results in about 2.2mW. By using the TOP current as opposed to the average current I am again over-estimating the power loss. 

We could use the much larger numbers in the datasheet as the turn-on-time and turn-off-time, but from the test-circuit in the datasheet I think they are driving the gate with a 50 Ohm impedance source. Thus slower times than what I just calculated are to be expected. If we do take those values, we get similar numbers to what I see on the scope (I suspect: my scope is too slow: I ordered a faster one). In  that case, the switching time is 10ns as opposed to 2.13 ns. Still power loss levels are much higher in practise than calculated. 

It turns out that my switching frequency was higher than I thought it was. I'm now at least getting numbers in the same order-of-magnitude than what I'm measuring. I've reduced the switching frequency by 2x and this seems to improve efficiency.  I'm now at around 89%. 

My calkculations of the switching losses was 1) using a lower frequncy than I was actually using. 2) missed 1 of the two switching moments. Now I'm getting results that are within an order of magnitude of the measured values. 

"Just measure" is not useful. The thermal camera measures "lots of power in the mosfet" I can throw the moseft away, buy new, different ones and see if that helps. That's useless as I don't know what parameters to pay attention to if the calculations don't show which parameter influences a signfiicant amount of power. 

Getting reasonable results from the calculations proved that the switching losses are difficult to improve upon except by reducing the switching frequency. The mosfets are "kind of small" compared to the usual LM3409 use-case. Thus the switching times are really fast, but switching more than a million times pre second adds up. 

In the end, I had to reduce switching frequency. My mosfets are now about 50C after a reasonable on time. 
One of the things that now contributes is the 2mA  standby current and the charges for the gates. For 16 drivers at a time I can consider to make a DCDC that provides VIN-6.5V. to the LM3409's. (Fun excercise because it will have to SINK current!)