I was interested in what the actual relative efficiencies were for dimming a FET driver with PWM vs using linear regulators (7135s), so I did some tests. I estimated output of an eagle eye X6 with XPL HI V2 3B using ceiling bounce. The drivers I tested were a FET-only driver from mtnelectronics, and a H17F driver that has a 8x7135 channel and a 1x7135 channel. In all cases, to measure the current, I charged the 30Q cell, used it in the flashlight for a set amount of time then recharged the cell to measure the charge used in that time and used these numbers to calculate an average current. This method gets around the potential inaccuracies of measuring current that is being pulsed with PWM.
In the tables below, the first column is the measured current. The second column is the lux I measured in my ceiling bounce test, and the third column is proportional to an efficiency of sorts; the lux divided by the current. For the FET driver I measured 100, 50, 15, and 5 modes. For the H17F, I measured the 8x7135 mode and the 1x7135 mode.
FET:
Ilux__eff.
5.88A96__16.3
2.58A50__19.4
0.72A14.4_20.0
0.22A4.5__20.5
7135s:
Ilux__eff.
2.91A66_22.7
0.37A10.7__28.9
So the 7135s are indeed more efficient than using PWM on the FET driver. In the 3A range, using the FET driver with PWM uses ~17% more power. In the 0.37A range, the FET driver setup uses ~42% more power. These numbers are significant, but the difference is not as great as the ~64% quoted in the LD-2 thread. I think the difference comes from the fact that they assumed the output of the PWM-dimmed light drops linearly with duty cycle. My tests show it drops more slowly. For example, when the current drops from 5.88A to 2.58A, a factor of 2.28, the output drops only by a factor of 1.92. The drop in output is less steep probably because the LED junction is less hot at 2.58A than at 5.88A.