If you can get over the MicroSlop bashing(opps, I did it also... sorry)
and the SGI hype, this talks about SMP a bit: from
http://reality.sgi.com/ariel/sgi-myths.html
++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Yet another gross oversimplification and a misleading statement often
expressed by reporters who have very little or no understanding of what
multiprocessing is, and what are the bottlenecks in a computer system
with more than one CPU.

         Reality:

              Having N processors in a system is a long way away from
having the system utilize all their power. Achieving this is mainly an
operating system (software)
              issue. It is much easier to make a system with N
processors (and there are several PC vendors stuffing their boxes with
additional CPUs) than having a system
              that actually works N times faster, or close to that, by
adding more processors. If you try one of these PC based systems, and
expect doubling or tripling in
              speed, chances are you're in for a disappointment. In fact
the current Intel's SHV motherboards are known to saturate their system
bus fairly easily with
              commercial workloads (like databases) at a low count of 4
CPUs due to low memory bandwidth and relatively small (512KB) caches.

              While Microsoft is working on scaling NT to provide good
support for 4 CPUs, SGI has working systems at customer sites supporting
128 CPUs, others
              supporting over 200 CPUs, and is working on scaling its
high-end to thousands of CPUs for general application loads. Meanwhile,
Cray has been delivering
              more specialized multi-processor systems since 1983 and
has shipped systems with over 2000 processors for specific application
mixes.

              In symmetric multiprocessing (SMP), the bottleneck pretty
quickly becomes something other than the CPUs. Even if we assume the
operating system issue is
              solved by making the OS kernel multi-threaded and adding
fine grained locks on shared data structures (which is a difficult
problem,) we are still left with
              all the hardware components that do not scale when we add
CPUs: notably the bus and the memory which are shared among all CPUs.
Pretty soon adding
              more processors will not speed the system at all, given
that the main bottlenecks reside elsewhere. This is why SGI is evolving
away from SMP into the new
              paradigm of cc-NUMA (Cache-coherent Non Uniform Memory
Access), and using a new OS foundation (Cellular IRIX TM) in its new
Origin TM product
              line.

         The point here is that SGI's/Cray's lead over the PC world in
multiprocessing and scalable systems is clearly expanding rather than
shrinking.