Titanio said:b) are running on a card with less bandwidth to memory than RSX. I'm not saying that those necessarily balance out the first flaw, just that it's not a one-way street.
I think this second point is not realistic at all.
RSX has a bandwidth of 22.4GB/s to the 128MB of GDDR3. To the XDR, RSX has bandwidth of 20GB/s write and 15GB/s read. So in a best case situation RSX has either 42.4GB/s of write bandwidth or 37.4GB of read bandwidth.
To compare, a stock 7800GTX has 38.4GB/s of bandwidth.
So in best case scenario the RSX will have 4GB/s more bandwidth, and in worse case scenarios (where RSX is given bandwidth priority over CELL) it will have 1GB/s less bandwidth. Without knowing the overhead of accessing XDR over a dedicated pool I think it is a little early to say definitively that RSX has more bandwidth than G70.
Furthermore, in the above example of maximum bandwidth (42GB/s) we are assuming RSX is eating all but 5GB/s of system bandwidth.
No one can say with a straight face that game designers are going to do that; and that just dovetails with the previous point: If you give CELL 5GB/s of bandwidth for system memory you are going to be CPU limited.
And with 5GB/s for CELL (which is going to need the bandwidth to get close to theoretical performance) you have less than a modern PC (6.4GB/s with DDR400). I don't think I need to connect the dots on how such a game would most likely look. The fact it is getting free AA because it is CPU bound probably wouldn't impress many gamers
In reality, the G70 will have quite a bit more memory bandwidth than RSX in most games.