I am currently waiting for my gp2x to get to me (tomorrow, yay!) so I haven't even begun to poke around or even port some of my own stuff to it, so take my comments on speed with a pinch of salt.
I really wouldn't bank on pygame being a smart choice for making complete games on the gp2x. (Unless 1-2 fps isn't a problem)
On my PIII machine at work, python/pygame's limiting, and pretty-much unavoidable step is the blitting function, which is expensive. I am thinking about re-writing some of the core pygame functionality directly into SDL & C++, for the added speed and ease of porting from pygame to C++, (but I think someone has beaten me to this). Pygame is a really good prototyping base though, as it lets you iron out the problem areas of a game before commiting it to a lower level language in which key problems may not be so obvious. (Ignoring low-level implementation differences, anyway)
Devs, or anyone with access to a gp2x: Is a game, running in python, accessing SDL through python bindings, going to run fast enough to be playable?