Writing a pure OpenGL engine is always going to yield better performance (assuming you’re an expert on what you’re doing). However, writing an engine for a non trivial game is difficult to get right (or good).
If you’re just playing around (hobby project) its really a decision you don’t have to worry about. Just use what seems most interesting.
From a commerical point of view, it would be far more cost effective to use an engine like Xith3D rather than go for pure opengl. You’ll just end up duplicating effort and not finding your bugs as quick (since not as many people will be using the code). Performance wise Xith (and other engines) are fine unless you’re trying to push the boundaries of technology.
You larger concern should probably be download footprint (que Cas), since most of the engines, including Xith, get quite large.
I’ve attempted a bunch of stuff with OpenGL and Java3D. Generally the performance of pure OpenGL looks really good up front but comes down quickly as the scene gets more complex. The tail off with the engines is slow but performance starts at a lower level.
Long post over,
Kev