[HH] Adapteva's Epiphany

[Bill Bogstad]

On Thu, Jan 26, 2012 at 1:42 PM, Tom Metro <tmetro+hhacking at gmail.com> wrote:

> Which include smartphones and tablets. Seems a bit overkill for those,
> given that the software hasn't yet caught up to really taking advantage
> of 4-core CPUs.

I tend to agree.   Most general purpose apps are not well optimized
for lots of parallelism.   In many cases, they are inherently serial.
 I just don't see
developers spending a lot of time updating their $1.99 app to use 16
cores for what is likely to be a very niche hardware platform.

> At "only" 1 GHz per CPU I don't think "supercomputer" is a compelling
> option.
>
> The strength seems to be in applications that need lots of parallel
> signal processing, like VoIP, video surveillance motion detection
> software video compression, etc.

If you are doing any kind of simulation over a 3D space (weather,
protein folding, etc.), I can see trading core speed for # of cores.
 You just change the granularity at which you assign individual cores
to be responsible for physical regions/objects/etc.   I think a lot of
supercomputing still fits into this category.
If I were them I would start dealing with the various manufacturers
who sell specialized add-on boards to soup up PC servers for
scientific computing.   It sounds like you could easily put 4 chips
(64 cores) onto a single board PCI express board.   Linux Journal used
to have ads from those companies all the time.

Bill Bogstad