Getting to the core of the issue

Three things:

1) The HPC crowd, especially those involved in anything fluiddynamik, need all the cores they can get. Our software is already MPI-ed up to the eyeballs; just give us the cores and memory, and we will use them.

2) Simplifying parallel programming. MIMD programming models are fearsome beasts, and horrendous to debug: you can spend weeks rather than days getting the buggering programs to work. But does anyone remember the fit of SIMD parallelism of the early 90s? The Connection Machine was actually /easier/ to program than the serial equivalent. (NVidia's CUDA provides an interesting avenue worth exploring)

3) To Martin: your models couldn't have been doing anything clever. Running state-of-the-art CFD models simulating anything larger than 10^6 m^3 requires a serious amount of horsepower. That's why they practically all run off the back of MPI/OpenMP.