Originally Posted by Ryeyoo
Well, it's 512 shared which does add more flexibility but I'm not sure it's taht significant to 256 for system and 256 for video. I'd imagine you could program it so that the ram streams from system to gpu relatively easily. I believe the 360s 10mb of dedicated gpu(digitalfoundry mentions it frequently) ram helps the 360 a lot more than the unified architecture.
RSX can read and write to both memory pools, altough with some restrictions in block size and so on.
As can Cell, but it can only read from GDDR3 with speed of 16MB/s. (basically it's for debug.)
The 10MB on X360 GPU actually resides on completely different die which also hosts ROPs.
This also means that only read and blend frame buffer operations can use the massive bandwidth it offers.
ALUs or texture units cannot do anything in it, but work solely trough main ram.
One of the big advantages is that blending is free in terms of bandwidth and that currently used framebuffer doesn't take any memory and MSAA is resolved when copied to main ram. (with 720p 4xMSAA this is around 28MB saved main memory.)