Probably no back-and-forth. The ideal way (i.e. minimal bandwidth) would be to transfer the framebuffer(s) from GDDR to SPE local store (in several chunks) and spit them back out to GDDR so it can be scanned out.I'm kind of curious to know how they do post processing on the SPU. I'm sure this is answered somewhere. Pardon my weak understanding of the graphics pipeline, but post-processing is done after rasterization, correct? Wouldn't that mean copying data back and forth from VRAM to main memory, since them segmented?
Post processing can refer to any screen-space operation that takes place after all objects have been rendered, i.e. after the frame is "complete", but before it is presentable. Tone mapping is the traditional example, but depth-of-field effects, MRT integration (deferred lighting schemes, certain types of motion blur) and aa resovle can all be shoved under that same umbrella.