Talk about bullshit. He claims that it is CELL+RSX which is capable of 51 billion dot products/s, when in fact, it is RSX alone, and this is clear from Nvidia's presentation.
And I love the "278Gb/s vs 22Gb/s", which is wrong on two counts. First, because the RSX can access both GDDR and XDR memory. And second, because of the phony bandwidth numbers, comparing effective with real instead of effective vs effective, and not looking at the situation holistically, since there is more to bandwidth than AA-framebuffer bandwidth.