SV: Capacity of the brain

Kragen Sitaker kragen at dnaco.net
Fri Sep 17 10:23:28 EST 1999


In article <bf8$WAAoxW43Ewnx at xerez.demon.co.uk>,
Shez  <newsreply at xerez.demon.co.uk> wrote:
>Extrapolating this to the number of neurones in the whole brain implied
>the power of about a thousand pentiums, or perhaps one decent sized
>supercomputer. (The pentium figure discounts networking overhead - in
>practice I suspect you'd probably have to network more like 4000
>pentiums to get 1000 times the power of one pentium.)

It's not that simple.  Some things parallelize much better than others,
with much less data being exchanged.  There's Amdahl's Law, of course,
but there are also coupling limits.

The Tera (http://www.tera.com/) folks are working on coupling limits.
They claim that distributed-memory processors like the T3E work around
5% of capacity on some of the codes they want to run well, due to cache
thrashing and communication.

On the other hand, some algorithms parallelize extremely well easily,
even with quite minimal communication.  There are a bunch of folks
called genetic-programming.com who are running a 1000-Pentium cluster
to try to invent algorithms so they can patent them.
-- 
<kragen at pobox.com>       Kragen Sitaker     <http://www.pobox.com/~kragen/>
Fri Sep 17 1999
52 days until the Internet stock bubble bursts on Monday, 1999-11-08.
<URL:http://www.pobox.com/~kragen/bubble.html>



More information about the Neur-sci mailing list