I just read the Dynotest article. The electrical load used for the test was a 12 ohm resistor with parallel 7.8 volt zeners. On the AXA power curve I see the knee in the curve occurs around 4 watts output, which corresponds to about 7 volts AC (rms) delivered to the load. The peak voltage then was 9.8 volts, so there should have been considerable Zener conduction, potentially the major source of non-linearity. For the less-capable gens knee powers varied from about 2.5 watts to about 3.9 watts (identifying the knee is not an exact science). These correspond to rms voltages from 5.5 volts to 6.8 volts, giving peak voltages from 7.8 volts to 9.7 volts. The zener was in some level of conduction for all of these generators, but this does not explain why the knees are not at the same power level. If the zener was the only cause of the knees the limiting power level should be the same for all generators. Something else is going on at the same time. I can only guess what it is, but my best guess is magnetic saturation.

If you want more than 3 watts from a generator and more output as you go faster, I'd suggest using a Schmidt, an AXA-HR, or any of the ones that tested higher than the Schmidt. I had trouble identifying the Shimano curves, but I think the power curve of the Shimano Inter-L overlaid mostly with the Schmidt. However, I don't know if that corresponds to any current Shimano model. It's hard to say "get Shimano" because I don't know if the current ones will match this good performance.

It would be real interesting to see how the better ones performed with a 10 ohm load, and with the Zener diodes removed.

I'm sure there's some good information at Candlepower as well, but I'm having trouble with my membership there.