From: Eugene Leitl (Eugene.Leitl@lrz.uni-muenchen.de)
Date: Thu Aug 16 2001 - 10:01:41 MDT
On Thu, 16 Aug 2001, Charles Hixson wrote:
> Just a few data points.
> 1) Yesterday I reread Staring into the Singularity, and did a bit of
So, has the Singularity started to stare into you yet?
> arithmetic. Assuming that in March the computer speed was, indeed,
> doubling every 18 mos. And assuming, as indicated, that each successive
Assuming, that pigs would fly, would a Glocester Old Spot strapped to a
hybrid scramjet/hydrogen/LOX propulsion unit go orbital?
> doubling takes half the time of the previous one, then in March 2004 the
> curve goes vertical.
That's some interesting algebra you're using here.
> 2) Yesterday in the news there was released an announcement of a 2GHz
> chip from a company that I had never heard of.
Them multiple gigahertzen is sure nice, but what's the memory optimum vs.
pessimum of that puppy?
> 3) I'm told that for not too much one can fit a modern PC with a board
> that supports 8 way processing.
Given that sentient stuff will be memory bandwidth starved, I don't see
how sticking more processors onto a single bus helps the issue. If
anything, one would want to farm hybrid good OPS/Watt ARM/DSP cores with
embedded memory, and packet-switched ultrafast serial networks.
> 4) This morning I read in an old InfoWorld (2 weeks ago?) that IEEE
> 1594b products will be released this fall, and that one implementation
> of it (using fiber optic cables) permits a 3.2 Gb (that's bit)
> connection over distances of 100 M. These are expected to be boards
> that can be fitted into an ordinary PC.
Not so ordinary, given that even a single GBit NIC is going to choke your
32 bit PCI bus, requiring 64 bit buses. It would be sure nice to be able
to cluster motherboards directly, using HyperTransport.
> 5) Beowulf clustering is now becoming common.
Sure, but what is the next step after that?
> 6) I think it was in C Friendly AI that the projection was made that a
> minimum system needed by a seed AI would be 32 processors running at 2 GHz.
Why not 42 processors? Or 420, or 2783144711?
> 7) So by the end of the year it should be possible to put together a
> system about twice the minimum strength for around $10,000.
Would make for a nice distributed game server.
> 8) So by the early part of next year it should be possible to put
> together a rack-mount system containing 64 1U boxes linked with cheap
> hot-pluggagle connectors that are interlinked at high speed (ehternet or
> IEEE 1594b, depending on what's available) with each node at least a
> 2GHz cpu (though the rack mount system would drive the price up a bit).
> That fits in a one wide rack. For less than, O, $20,000.
Suggestion: use Myrinet (or channel bonded FastEthernet, HP switches),
off-shelf (maybe dual) Athlon DDR CL2 boxes, cheap metal shelves, and
invest the rest into a really good airconditioning unit (unless you live
in colder climes, where you can just open a window, assuming you're
unafraid of enterprising polar bears).
> Projection 8 can be done with currently available on the shelf hardware,
> except for the cpu speed. The new stuff just makes it cheaper and
> faster (and slower, if you need to wait to get it and debug the setup).
>
> 9) These rack mounts could easily be connected with a fast interchange.
> Take one computer from each rack and devote it to managing task and
> information sharing between and withing racks. You could fit an awful
> lot of racks along 100M of wall space. And since they are lined up
> along a wall, they'd still all be easy to get into to fix.
>
> As always, software is the sticky point. Unless someone has some
> intelligent code writing software...
Oh yeah, you can clog up very fat pipes with bloated protocols, and
parsing XML the OO way will make you realize why you invested in so much
hardware which will become obsolete before the year has grown very old.
> 10) The assertion has been made that the more computing power you have,
> the easier it is to end up with an AI (though not necessarily a friendly
> one). So by the middle of next year any reasonably profitable business
> should be able to put together enough computing power, if it chooses to.
It is definitely easier, though there is a threshold. If the Moon wasn't
made from green cheese but computronium, it would take a whole lot of hard
work before it started to look down to Earth as an easy snack.
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 08:09:52 MST