[rescue] Cray J90s
rescue at sunhelp.org
Tue Jun 12 17:29:47 CDT 2001
On Tue, 12 Jun 2001, Dave McGuire wrote:
> Another good point to ponder...some things vectorize better than
> they parallelize.
Neural nets are a terrific example. Parallelizing each node kills you
with overhead, but SIMD vectorization is a huge win. You can still use
parallelism to anneal and run multiple training passes simultaneously but
this is iffy at best for speeding up the process.
Some things don't vectorize or parallelize at all. These are the worst.
I still have my VIS enabled backprop neural net library, it would be
interesting to see how much of a benefit could be realized by vectorizing
the whole thing. Ran like a bat out of hell on a cluster of E450's.
More information about the rescue