[extropy-chat] Optimal computer configurations [Was: Commentary: Does Karl Schroeder's opinion reallymatter?]

[Robert Bradbury]

On 4/29/06, Russell Wallace <russell.wallace at gmail.com> wrote:
(regarding pigging out on memory because it happens to be "cheap")...

> Good, good, good!
>

Not really!  Hypertransport &  multi-gigabit ethernet = good good good.
Cell microprocessors are kinda good good good.
Highly parallel graphics cards are somewhat less good good good.

Most current systems do *zippo* with respect to making nanotechnology
development faster.  And similarly contribute little to possible development
of an AI.  If you really want to do high quality atomic scale simulations
you need processor and memory architectures that fit the basic requirements
for molecular modeling.  Those are *not* for cheap memory widely separated
from the CPU(s).  You want lots of processors with good F.P. capabilities
with access to their own local memory pools.  The Cell is the closest thing
to that currently.  You also need high bandwidth between as many of those
processors as you can afford.  That is where hypertransport & multi-gigabit
ethernet come in.

If you look at the AI side of the equation, you have to consider the brain
is a physically concentrated of relatively limited capability processors
with very high aggregate bandwidth.  We are moving towards this with
multi-cores but the current primary PC architecture (huge separation between
CPUs & memory) doesn't even come close to what is required.

A significant limiting factor on continued progress in computer hardware is
> demand going down because too much programming effort is spent wasting
> computer capacity (by leaving it lying idle) rather than using it to improve
> reliability (for a start, by switching to languages other than super macro
> assembler! :P), functionality and usability.
>

Reliability is improved by redundancy -- you are starting to see that with
web applications (my computer can fail but I can still go to any other web
connected computer and read Gmail).  Usability is what is driving things
like myspace & youtube -- not that they are contributing in any significant
way to nanotechnology or AI.  Functionality is an interesting topic.  It
took us ~20 years to go from C to Perl and another decade to get to Python
and Java.  And though I don't claim to know the last two my limited
awareness doesn't point out significant differences between them and C.
(Yes one doesn't have to handle memory allocation but of course that can
lead to memory fragmentation which leads to the problems one can currently
encounter in Firefox.)

Serious workloads like simulations always need more computing power, but the
> people running them don't have the money to pay for chip factories at
> several billion a pop. It all comes down to the people writing programs like
> Firefox and Doom 3 to put the power to mass use - let them be praised, not
> criticized.
>

Hmmm... Go ahead and make the case that Firefox is contributing to computer
architecture development will support cheap simulations...  I doubt it can
be done.  Anything you suggest that Firefox is doing driving the limits of
the hardware I would suggest may be an unconscious and unnecesary waste of
resources (I haven't heard about people complaining about Opera being so
problematic).  You can make a better case for games driving parallelism and
even cooling technologies -- but I don't believe nVideo or ATI have released
the specifications for their GPUs for general purpose programming (the
drivers are still proprietary AFAIK).  So all you can use the cards for is
displaying pretty pictures.  (This stands in contrast to IBM pushing the
Cell as an open platform.)  I agree that there is a funding problem for what
we really need -- IBM could have easily produced a chip optimal for
molecular modeling 5+ years ago.  They were well aware of what was required
(there are papers in the IBM Journal discussing it).  However it took
teaming up with Sony and a potential market of 100 million chips (e.g. the
PS3) to get them to produce something even halfway close to what is
required.

Robert
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20060430/f9fda6f1/attachment.html>