Finally ...... Japan realising that in Order to get things done right, one has to be able to do everything on one's own, although that does not necessarily mean that one has to do IT alone. Bravo.

And in SPARC HPC AIR&dD Circles, there are Complementary Parallel Fields which Increase Operating Systems Penetrations Testing and MetaData Acquisition and Spontaneous Production/Creation Rates, Exponentially, which Leads to Rapid Dominant SMARTer Control with PreProcessed, Chain Linked Power Supplies aka Source Lode Generators** . .....from and for NEUKlearer HyperRadioProActive IT and Intellectual Property Power Units.

* For Virtual SuperPowers.

** New Fangled Entangling Lead Strings for Thread Weaving/Novel Application.

There are probably a number of technical issues we don't get to hear about that will also come into play. The grunt per unit power of the new Sparc will be interesting, and may make it a clear winner over an x64. SGI's ccNUMA machines may be a hard habit to break depending upon the simulation paradigms used. For many real codes the efficiency of coding in a single shared address space can mean the machine is much faster than a clustered system, even if both have similar Linpack benchmarks. Vastly nicer to code in too.

But scalability isn't easy at these sizes, and it is likely that simple porting of codes is the least of their problems, as the relative patterns of data, communication, and compute can shift significantly as the fidelity of the simulation changes. This may lead to significant structural changes in the simulation simply to avoid quite awful slowdowns in the code. In principle, it is these patterns that should drive the choice of architecture.

But, as noted in the article, at this level, politics plays a very big part in the answer.