[Beowulf] Pony: not yours.

[Eugen Leitl]

On Thu, May 16, 2013 at 01:37:05PM -0400, Mark Hahn wrote:
> >On Thu, May 16, 2013 at 10:28:12AM -0400, Mark Hahn wrote:
> >>- do we need exascale anyway?  would the world be better off with a thousand
> >
> >Yes, EFlops is entry level for projects like
> >http://www.wired.com/wiredscience/2013/05/neurologist-markam-human-brain/all/
> >and if you want invididually accurate neural emulation
> 
> sure - such suggestions are pretty easy to put together, but they do
> not answer the question: should it be done, and at what cost?  I

Working real AI or personal immortality are quite useful, at least
for some people, I could imagine.

> still think
> they're really being used as a fig-leaf to cover the erection people
> get when imagining the "next-a-scale" facility.

I dislike big systems, and be it only becase they're resource
intensive (and hence not exactly egalitarian computing). 
Unfortunately, a lot of very practical problems require quite 
massive resources, and preferably under everybody's desk
who would care to get one.
 
> simulations like this are extremely dubious in that they're fundamentally
> limited by the quality of raw configuration data.  it is

If you look at holistic approaches like http://www.openworm.org/
and http://www.si-elegans.eu/ these are quite easy to validate,
given a sufficiently diverse behaviour library. Markram is more
basic research in comparison, to obtain the parameters for 
the biologically accurate simulations.

> emphatically not clear that neuroscience is primarily limited by the
> scale of machines to run simulations.

It is pretty much limited, if your primary input is the annotated
connectome, at few nm resolution.
 
> >>power-efficient systems?  if power is an important TCO component, why aren't
> >>we optimizing for it already (in any sized facility)?
> >
> >I definitely do for my private projects, as 0.25 EUR/kWh are
> >a pretty good argument.
> 
> even retail, fully-delivered rates here are half that.  but out of curiosity,

Well, at least it's 100% fully renewable juice.

> does your comment mean that you, for instance, buy arm clusters, or
> at least LV versions of x86 chips (and high-core, low-clock models),

Yes, both for my personal computing (mostly, just a few old boxes on
the Internet, less than kW total) and the dayjob I do that. I would 
gladly buy ARM cluster-in-a-rackmount, provided the price/performance 
is right, and assuming you could get these from the usual vendors.

> gold/platinum
> PSUs, etc?  (or just do your compute in a location with cheaper power?)

Would love to crunch interesting systems for a living, but that doesn't
get the bills paid. I hope to get something done with the Epiphany, assuming
they keep the price point and availablity (and also the power envelope).
 
> regards, mark hahn.