Owen Densmore wrote:
It will be nice when we can try HPC easily at home, so to speak.
AMD's OpenCL toolkit scales well on Opteron system. Also, Snow Leopard
comes with an OpenCL implementation and scales well for small problems
on the GT9400 and GT9600 found in a MacBook Pro.
http://developer.amd.com/gpu/atistreamsdk/pages/default.aspx
From a educational point of view, the problems of GPUs (high latency)
are a benefit because an experimenter has to learn how to deal with an
`unbalanced' design. In the future, more and more communication will
be on chip, but asynchronous. Finding ways to deal with that will be a
challenge, and the people that have the skills to do it will be at an
advantage in the HPC market.
There is a lot of useful investment that can be done at a small scale,
and doesn't require having access to tens of thousands of boxes and
fancy interconnects. The hard issues for coarse-grained communication
will be related to resilience, when nodes go down or give bad results --
in large part figuring out algorithms for adapting to a compute fabric
that is not fixed. Even researchers that do have access to such
fabrics don't have direct control over them. A given, more or less,
and you deal with it.
http://www.nvidia.com/object/cuda_opencl_new.html
http://www.alphaworks.ibm.com/tech/opencl/faq
http://www.pgroup.com/support/downloads.php
Marcus
============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
lectures, archives, unsubscribe, maps at http://www.friam.org
