I was wondering which hardware supports direct transfers from the frame buffer to texture memory, without a detour via CPU memory, in particular for OpenGL''s glCopyTexSubImage2D. How do the OpenGL implementations handle this?

For glCopyTexSubImage2D this depends on the driver for the gfx card.

For Radeon: Not accelerated unless you use a pbuffer.
For GeForce: If the framebuffer''s pixelformat is the same as
the texture''s pixelformat, it''s quick, if they are not the same it seems to converted by the CPU (slow).

Coder10000