About the HD 2900 5 instructions per clock... I remember that when I saw all the specs and diagrams, my guess was that despite AMD's claims it was more likely that the chip wouldn't be able to effectively use both scalar and vectorial units at the same time. Can't remember why though, and from that day until now I look at the HD SPs as 4 ops/cycle units.
I'm too lazy to go back on reading all the stuff, so I would love your opinion. So please everybody share your thoughts on this.