[ExI] gimps

[Adrian Tymes]

On Fri, Jun 22, 2012 at 7:59 AM, spike <spike66 at att.net> wrote:
> So we have gained a prefix in 14
> years.

Which about tracks with Moore's law (doubling every 1.5
years -> gain a prefix every 15).

On Fri, Jun 22, 2012 at 8:55 AM, Eugen Leitl <eugen at leitl.org> wrote:
> The clocks effectively stopped doubling in 2003, and it's been not so
> good in general http://www.gotw.ca/publications/concurrency-ddj.htm

Yeah, yeah, transistor growth has gone into parallel processing
more than clock speed increment.

> In order to continue scaling as long as Moore's going to hold (not
> for very much longer, the 2d litho limit is probably around 1-3 nm)
> we need different architectures, and people who can deal with asynchronous
> shared-nothing distributed (and even nondeterministic) applications.

Already happening.  For instance, games w/high end graphics farm
out their rendering tasks to different cores where possible.  Often,
this is done via software libraries, so most of any given game's
developers don't have to worry about it.  And this happens on a
computer which is often running other tasks (like a VOIP client to
connect the gamers) at the same time.

Or, for non-games, consider Web and database servers for large
businesses.  How many threads do they run in parallel on most
days?  As many as they can.  (Ever seen a server try to deal with
about 10,000 simultaneous user requests?  I have, and we'd only
expected a few hundred simultaneous peak users.  The system
held up, but we rearchitected for bigger peaks once the event was
over.  This wasn't even a large business; this was a 5-man startup.)

And then, of course, there are scientific applications like Spike's.