[Beowulf] OT? GPU accelerators for finite difference time domain

[Joe Landman]

Mark Hahn wrote:

>> Has anyone any experience of this? How do these products stack
>> up against the traditional Beowulf solution?
> 
> they _are_ in the spirit of Beowulf, which is all about hacking
> commodity hardware to suit HPC purposes.

... but it would be hard to fit into a 1U case, and the 200+ watt power
requirements could be daunting to smaller supplies.

Not that I am against GPUs as accelerators, on the contrary.  Just be
aware that GPUs with significant calculation capability also will
require a rather significant power supply and cooling airflow.

Right now, accelerated computing is in its infancy.  You have host based
(SSE*), and host attached  GPUs, FPGAs, APUs (Accelerator Processor
Units) in general such as ClearSpeed et al.  You can think of clusters
as accelerators in the sense that they provide a larger number of cycles
per unit time to your application.

There are no single APIs to bind them all though.  A number of the APU
people are realizing that there is significant benefit to providing
acceleration behind existing popular interfaces, as it lowers the
barrier to adoption and usage.  If your code is designed with FFTW in
mind and you have to re-organize your arrays to suit another FFT
implementation, this can be both annoying for the programmer, and
inefficient.

Regardless of the accelerator you choose, expect *some* rewriting of
code at minimum.  Current GPUs are focused upon singles and ints.  As
Jeff noted, doubles should be coming.  As Mark noted, slow doubles
aren't useful.

-- 
Joseph Landman, Ph.D
Founder and CEO
Scalable Informatics LLC,
email: landman at scalableinformatics.com
web  : http://www.scalableinformatics.com
phone: +1 734 786 8423
fax  : +1 734 786 8452 or +1 866 888 3112
cell : +1 734 612 4615