Computer power can grow in three ways: faster CPUs, more CPUs in a box, or more boxes tied together. SMP, or symmetric multi-processor, systems have been the forefront of high-end server technology for the past two decades. The bottleneck of the technology is sharing memory among multiple processors via various caching and shared-bus systems. NUMA, the option of choice, scales to about 64 processes, but as of yet, not further.
The article discusses some of the particulars of Sun's COMA implementation, apparently shared between both hardware and software. COMA has been proven academically, but not yet used in commercial products. There are issues both in system and software design which may make or break its success.
I'd be interested in hearing from the EE wonks as to what the up- and down-sides of this design are, and (of course) any thoughts on whether Linux has any potential to move into higher-end SMP space with either NUMA or COMA designs.
For some background reading, the following EE Times article, BM, Sun eye NUMA architectures to make servers sizzle, may be of interest. Also recommended is Greg Pfister's In Search of Clusters, a highly readable book on high-performance computing.