Didn't you write on rwt just before K8 arrived that you thought that K8 would have a 10-15% performance improvement due to the integrated MC? I.e. about half of the total improvement you expected? I remember you said something like this (might be off slightly) because you turned out to be quite right.
I don't remember. I might have made an estimate based on K7's performance scaling with frequency on SPECint2k and a guess of how much memory latency would be saved.
The problem is the proportion of the performance difference between K7 and K8 attributable to the integrated memory controller varies wildly from program to program. At one extreme are low ILP codes dominated by serial data depend- encies (e.g. pointer chasing codes) in which the IMC gives nearly 100% of the improvement. The other extreme is very cache friendly code in which the IMC gives basically 0% of the improvement. The "average value" is highly sensitive to the selection space of programs considered as typical or representative.