X-Message-Number: 26325
From: 
Date: Mon, 13 Jun 2005 01:13:56 EDT
Subject: Uploading technology (1.iii.0) What technologies ?

Uploading technology (1.iii.0) What technologies ?
 
The Blue Brain Project and the CCortex project [ http://www.ad.com/tech.html 
] both, look at building a brain on a computer. The CCortex seems to 

concentrate on "classical" neural networks, so it would be a brain scale neural 
net 

with elementary neuron far more simple that the biological ones. The BBP is more
ambitious and could ultimately represent the computer part of the uploading 
system I have described before.
 

These projects are based on general purpose computers, the CCortex would use 
500 linked workstations and the BBP is based on a version of the Blue Gene 
computer from IBM. This is a multiprocessor system. Given some raw computing 

power, it has been demonstrated by experience that, with a good interconnecting
scheme such the hypercube, the more we divide the pover into smaller processor 

the best the overal performance. T. Donaldson has championed this idea with the
concept of one neuron on one processor. It must be pointed out neverthless 

that this is true only for general purpose computer. When a system is built for
a particular task, be brain simulation or astronomical "experiments", this is 
no more true. A single processor or a small number of them is best. Whatever 
the issue, that remains limited at the simulation part worked out on the 

computer. for fast processes, that technologiy is far too slow and other 
solutions 
must be sought.
 
 
As pointed before, FPGA seems the best contender, even if specialized parts 
must be worked out by TRACs, their analogic equivalent. General purpose 

computers with one or many processors are readily available, FPGA too and TRACs 
can 
be ordered as ASIC products with low mask cost. There is yet another 

possibility, the use of neurons on a chip. These are very special circuits built
to 

emulate biological neurons. Kwabena BOAHEN call them neuromorphic microchips in
the Scientific American May issue. The retina, may be the best understood part 
of the central nervous system, has been wired with its five layers, now the 
work is moving to the visual cortex six layers, the number IV has been worked 

out. [www.neuroengineering.upenn.edu/boahen ]The corresponding chips are not on
the market but could be ordered from the inventors for research purpose. Some 

organizations such MOSIS [www.mosis.org ] can handle the process of interfacing
with a silicon foundry for small batch, down to 50 items. 
 
One possibility is then to create a circuit on computer and FPGA and then to 
move it, or at least the FPGA part, to a TRAC-neuromorphic chip. Today works 
concentrate on simple neuron models, far from what is asked for in a uploading 
project. So it is not possible for that use to simply pass a phone call to a 
research lab a buy some chips added to their next batch order.
 
K. BOAHEN gives the following scale in processing power and efficiency :
The brain uses 10 Watts for 10^16 synapse events per second. The common 
computer today uses at least 100 W for One billion operations per second. The 
energy efficiency is 10 millions worst in the computer. But: Transistors are 

working 25,000 times faster that ion channels, if they worked at the same 
electric 
potential, they would be 3,600 faster than the neuron ion channels. This give 
the order of magnitude of the time sharing re-use of a given circuit.
 
Yvan Bozzonetti.


 Content-Type: text/html; charset="US-ASCII"

[ AUTOMATICALLY SKIPPING HTML ENCODING! ] 

Rate This Message: http://www.cryonet.org/cgi-bin/rate.cgi?msg=26325