Java Cool Dude did do some one on one comparisons with his demos way back and found Xith3D to be quite a lot faster.
What kind of demos are these. do they do ANY offscreen geometry? Do they contain a complex array of graphics states?
If not then I am sorry they ARE worthless microbenchmarks. At least as applied to any 3D environmental game. (Thy *might* have limtied applciation to doing 2D games in 3D but why yould use a scenegraph for that at all is abit beyond me.)
Same questions apply to your board game. I am perfectly willing to grant you that for very simple things a very simple API without some of J3Ds over-head might perform better. Thats not a general proof that "Java3D is slower" however, merely a proof that simple apps are better supported by a scene graph that does not have all the over-head needed to make complex things fast.
(The vast majority of modern games, it should be noted, fall into the "compelx" category.)
The reason I am so strong on this point is that every time I've actually looked at someones "proof" of this contention that Java3D is in general slower then some other scene graph, its always fallen apart on analysis. And I've looked at quite a few.
Do you have a pointer to them so I can see them run?
Edit: Also, have these benchmarks and test been run on a multi-cpu box? All the next gen consoles are multiple core. Thats available now on the desktop and I would expect we'll start seeing it stadnard in high end game rigs fairly soon. Another place you pay a small amount in J3D over-head is in support of the fact that is extremely parallizable and automatically handles multiple-cpus for you...