[extropy-chat] flash crowd super computer

[Eugen Leitl]

On Mon, Feb 23, 2004 at 12:51:28PM -0600, Damien Broderick wrote:

> Actually, it could be an idea from a Dam Clemmensen post to this forum, 
> cited in THE SPIKE seven years ago--except that Dan's approach to a home 
> brew singularity-grade AI *didn't* require all the machines to be 
> physically `in one place', which is a bizarrely 1940s' idea. But hey.

It isn't as bizarre as it sounds. If you want to be able to talk any machine
to any other without interfering, you need a lot of wire, and a big crossbar
(or equivalent) switch. For many reasons, chiefly because of minimizing fibre
length and because crossbars don't scale nor do long-haul protocols suit to
switching the Internet is a (meshed) tree.

If you bring many machines into one location, and make them talk pure Ethernet
frames instead of TCP/IP over arbitrary carrier you can achieve a far greater
bandwidth crossection (and a vastly smaller latency to boot) with lots less
thaler$.

Having this said, it's largely a synthetic benchmark. A lot of codes rely on
modern interconnect fabric such as Infiniband, Myrinet and SCI. These do make
even GBit Ethernet look very pale in comparison.

An AI to end all AIs won't need a globar crossbar, of course. It should be
able to live on a globally meshed fabric with FastEthernet port equivalent,
on a GBit/10 GBit backbone.

-- Eugen* Leitl <a href="http://leitl.org">leitl</a>
______________________________________________________________
ICBM: 48.07078, 11.61144            http://www.leitl.org
8B29F6BE: 099D 78BA 2FD3 B014 B08A  7779 75B0 2443 8B29 F6BE
http://moleculardevices.org         http://nanomachines.net
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 198 bytes
Desc: not available
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20040223/b1000b4e/attachment.bin>