-Right now, my computer (old powerbook OSX 10.4.11 - one processor, built-in graphics (ATY,RV360M11 it claims)) displaying at 1280 * 960, with 23 polygons being rendered at 30fps and nothing else being done in pd, claims to be using 13-17% of the CPU, but it doesn't seem to display correctly ('slow') + it becomes a bit unresponsive ('slow').

Make sure the -nrt flag is used with GEM. This should increase GUI response on OSX.

-A collaborators newer 'macbook' (not pro, but still 2 processors), with built-in intel GMA 950 gfx card, claims to be using 95-100%(completely incoherent), and is also unresponsive('slow'). Anyway, the instalation will probably run in linux desktops so i am trying to define the architecture a bit.

The Intel 'GPU' uses the CPU for just about everything. Performance is not good.

1- Which parts of this process (or image processing in general) are done by the CPU, and which by the GPU?

Anything pix_ is done on the CPU. The Geos use both CPU and GPU and the Manips and controls use the GPU.

2- shouldn't a two processor machine increase the performance in a GEM Patch by default: display in one and process in the other???

No. OpenGL commands have to be run in the same thread. Some of the video handling APIs (Quicktime and DirectShow) use multiple threads for encoding and decoding video.

3- How much and/or in what ways would the performance of the program increase if I had a better graphics card? What features should I be looking for in a card?

The ATI 2000 and Nvidia 8000 series should have little problem running patches with lots of geometry and basic GEM functions. Shaders can be performance problems even on the top cards. Large screen resolutions and fragment shaders are particularly hard on the GPU.