From: Yan King Yin (y.k.y@lycos.com)
Date: Sat Nov 29 2003 - 13:18:37 MST
--------- Original Message ---------
From: Eugen Leitl <eugen@leitl.org>
>I'm not complaining. We have MPI, nanokernels, architectures
>like Blue Gene (<http://moleculardevices.org/bluematter.pdf>
>demonstrates an application, and cites architecture papers),
>coarse-grain FPGA+embedded memory+signalling fabric as
>well as fine-grained cellular architectures. MRAM is only
>a couple of years remote from the marketplace, spintronics
>is soon to follow, molecular memory and cellular logic is
>10-20 years away.
There's 1 more question: Will the von Neumann
architecture be sufficient for running large-scale
neural networks, or is the bottleneck so severe
that a paradigm shift is required?
If fault tolerance can be easily built into memories,
with conventional logic operating on them, can this be
adequate?
Say a brain-equivalent NN will require about 150Tb of
memory (assume 1 synapse = 1 byte). And assume that
any address is connected to a fraction of this 150Tb
and has to be updated continually on the ms time
scale. We can estimate the bandwidth on the
data/address bus required for this amount of
processing and see if it can be met.
YKY
____________________________________________________________
Get advanced SPAM filtering on Webmail or POP Mail ... Get Lycos Mail!
http://login.mail.lycos.com/r/referral?aid=27005
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:43 MDT