Supercomputer cooling

http://www.geek.com/news/geeknews/20...1121022778.htm

On Fri, 21 Nov 2003 18:28:54 GMT, "Yousuf Khan"
wrote:

http://www.geek.com/news/geeknews/20...1121022778.htm


"A website called Queue is now arguing that supercomputers should be
judged on how much performance they can provide for a given amount of
energy/heat dissipation"

There is a problem with this approach that has to do with the
granularity of the parallelism. You can build servers or
supercomputers with lots of little chips running rather slowly, like
the server blades that use the low-votage Pentium M running at 1GHz.
No problem if you can break your problem into small enogh chunks that
a single Pentium M can make things happen fast enough, and if you need
to do things faster you use two Pentium M's.

That doesn't always work, or, more properly to say, it isn't easy to
get to work the way people are accustomed to programming. In any
case, you eventually run into Amdahl's law, which says that the
speedup you get from paralleization is no larger than the inverse the
the proportion of the code that is serial. A slight generalization of
that gives you a set of upper bounds N/T, where T is the propotion of
running time the code can only use N processors. The smallest of
those upper bounds is obviously an upper bound. That's why it isn't a
slam dump to conclude that fast, hot processors are an endangered
species for massively parallel calculations.

RM