He removed reliance on GC in the game-loop altogether _completely_. Then added and improved view-frustum culling, and usage of 16.16 fixed-point. Texture complexity was no issue, so went back to using crisp huge textures.
Framerate initially was like 13fps, now never falls under 33fps, is 50 at many places, and 75 at some places.
Info about framerate is sent via USB to Eclipse, and is consistent with observation (by my trained eyes ^^ ).
Float-to-string code was supplied, so that he could write framerate info to a char and send over USB.
Objects rendered are about 12. Qualcomm are not expected to release info on their JNI-using Neocore demo.
In my own code, I'm rendering all my geometry into dynamic vertex arrays, but limited only to floats so far - which is the limiting factor in my case (IntBuffer's fix will come with next FW update). Apart from fillrate limits when blending.