[Beowulf] Power calculations , double precision, ECC and power of APU's

[Michael Di Domenico]

On Tue, Mar 19, 2013 at 6:12 PM, Joe Landman
<landman at scalableinformatics.com> wrote:
> On 03/19/2013 06:06 PM, Michael Di Domenico wrote:
>> out of your many foot in mouth statements over the years this must be
>> the worst.  i can absolutely assure you there is an aspect of the HPC
>> community that does not care about DP, and cares more about SP
>> performance, and even more so about integer performance.
>
>
> +10 (on the caring about things other than DP, and even SP).
>
> There's a fairly large cross section of people whose simulation, codes,
> etc. have to deal with other-than-*P performance issues. Instruction
> graduation rate, IO and network bandwidth ... all of these things matter
> for a large and growing segment of the HPC community.  I've argued many
> times that basing the "fastest supercomputer" metric off a parallel LU
> decomposition manages to miss an ever larger segment of the HPC
> community, for whom that metric is not useful.

You're absolutely right.  While i absolutely respect HPL (and it's
creators) as a valid benchmark and we do use it is a generalization on
expected performance.  It doesn't give us any real indication on the
maximum potential of a computers performance with respect to our
codes.