Actually Fugger, SMP is often not a consideration at all for clusters. It all depends on your application, but if you need lots of memory, memory bandwidth, and network bandwidth per CPU, you never want more then 1 CPU per motherboard. A CFD problem fits this desciption very well due to the very large matrix sizes involved.
If you have smaller individual tasks, and stuff that can be efficiently threaded (as apposed to message passing), SMP is great and can significantly reduce the size & cost of a cluster. I may have to go with an Intel solution just because it can fit in one rack instead of two, despite the performance hit :-(
And by the way, trans-sonic CFD is NOT an "extreme light duty" application by any means.
My guess is Boeing went with AMD because they could get by with 96 CPUs instead of 120 Intel CPUs, and at a lower cost. In fact, this is exactly what Linux Networx reccomends for floating point compute bound applications. I know, because I was just discussing a 32 node cluster with them not 5 minutes ago. They promised to send some AMD vs. Intel floating point benchmarks with the quote.
In theory, there is no difference between theory and practice.
In practice, there is.