RE: Hardware Progress: $212/GFlop/s

From: Dani Eder (
Date: Wed May 07 2003 - 11:31:25 MDT

> > The processing power to run a human-level AI is
> > unknown, but is estimated to be in the range of
> > 100 - 100,000 TFlop/s.
> Where does this number come from?

The lower number is derived by Hans Moravec from
his work on machine vision. He took the processing
power to replicate the function of the retina, then
extrapolated to the entire brain by scaling the
mass of neurons in the retina to the entire brain.

My estimate is derived from: 10^11 neurons in
brain x 10^4 synapses/neuron x 100 Hz firing rate
= 10^17 synapse firings/sec. If 1 synapse firing
= 1 bit of data, and floating point calculation
= 32 bits, then brain is effectively 3,000 Gflops.

The range in the estimate is due to uncertainty in
how many bits of data a synapse firing is worth, and
what fraction of the human brain is needed for
intelligence vs. redundancy and autonomic functions.

Adding to the uncertainty is the level of clever
or poor design in an AI relative to the design of
the human brain. The thousandfold range is my
subjective estimate of the uncertainties.


Do you Yahoo!?
The New Yahoo! Search - Faster. Easier. Bingo.

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:42 MDT