As the market changes …

I’ve argued for a while that accelerators are going to be a creative and destructive force in HPC. They profoundly change the cost per cycle landscape, as well as the number of cycles per unit time. I’ve pointed out here that despite some misunderstanding of transformative technological trends, that better cheaper faster is one of the most important driving forces in HPC. It is a viable business model if you can figure out how to get the appropriate traction, and its very hard for larger organizations to adequately respond. That is, without simply buying the smaller company with better cheaper faster.
Michael Feldman makes a number of points in this regard. This is specifically about Westmere (the new Nehalem part), but he correctly generalizes to the rest of the industry.
A few points worth noting.

First, I agree that clusters will not go away. But the market will change, in what I have called a bi(tri?)furcation.
John West of InsideHPC asked readers to tell AMD what they could do with 48 cores in a single machine as part of a contest from AMD.
Think about this. Not so long ago, 48 cluster machines meant 24 physical servers. Then 12 servers. Now 1 machine.
A deskside at that.
Fermi is coming out with 512 cores. Lots of DP FP data capability.
Second, as noted above, the computational power will continued to be expressed/enabled via some sort of parallelism. Processors won’t grow much faster, just more numerous.
Which means aggregate performance and usable performance will matter. 4 operations per clock cycle aren’t all that relevant if you can’t ever get more than 1 operation per clock in your code. All that parallelism … how to use it? Is it well matched (bandwidth/latency) to the other subsystems?
If you hide all the performance behind bandwidth or latency walls, or make it hard to program, chances are it won’t be used in the market.
The short of it is that these new super-workstations will be in wider use for running day to day computing.
This will have the net effect of decreasing the workload on many computing clusters, and possibly decreasing the market for such systems. The large systems, less such a change, as people still need the big/heavy metal.
But I do expect that more computing will migrate to personal supers as a result of accelerators and APUs.
Moreover, this could also dramatically increase cloud adoption in HPC. Run it local, and only buy what you need to run it remote when you need it.
But in all cases, the market that gets destroyed is replaced by a much larger and much more accessible market. And this is good for HPC.