From: Phil Goetz (email@example.com)
Date: Fri Sep 09 2005 - 12:54:36 MDT
--- Phil Goetz <firstname.lastname@example.org> wrote:
> Making this mapping onto output states consistent, clear, and
> error-free requires a lot of engineering overkill. A neuron
> operates at a few millivolts, and its output, as a result, is
> "wrong" about 10% of the time. A silicon chip operates with
> about a 10-volt difference between its two states, in order to
> be almost error-free. We humans have gone from a reliability
> of .9 to 1-10^-14 or whatever - about a 10% improvement - in
> exchange for an increase in power requirements of 4 orders of
I spoke of power but gave numbers in volts. Sorry.
A head dissipates about 25 watts, and has about 10^10 neurons.
A CPU uses, what, 35 watts or so, and can do roughly one
instruction at a time, but each one is much faster, about
one billionth of a second, whereas a neuron operation is around
1/200th of a second.
Calculated this way, and discounting a factor of 100 for
neuron redundancy, the neuron comes in just about 1 order
of magnitude ahead of the CPU. Remarkable. A further
discounting of the fact that we're comparing single pulses
to 1/10th of a floating-point operations would probably
make them come in even. (A Pentium IV takes about 10 cycles
per floating-point operation, averaged over time so that this
is factoring in pipelining and such.)
Click here to donate to the Hurricane Katrina relief effort.
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:52 MDT