From: Dani Eder (danielravennest@yahoo.com)
Date: Thu Feb 07 2002 - 11:45:05 MST
--- Alfio Puglisi <puglisi@arcetri.astro.it> wrote:
> I just browsed Pricewatch, and I was able to put
> together a dual AMD
> system for with $340/Gflop/s. The resulting node is
> not balanced (not
> enough memory and storage space to match the
> GFlops), but is a
> faster number cruncher with slightly more memory.
>
Absent a real design for a human-level AI, I made
the following assumptions to come up with my
node characteristics:
(1) A human-level AI should have approximately
human-like response times. At 1 byte of memory
per 32 bit floating point operation per second
a node will be able to churn through it's memory
in 1/4 sec in theory, which is around the reaction
time of humans.
(2)A human-level AI will probably have approximately
the total complexity of a human brain. The human
brain has ~10^15 synapses. If the estimate of
100 Tflop/s for human-level processing is correct,
then 10^16 bytes storage allows for 10 bytes per
synapse to describe it's connectivity and state.
I recognise that 1 byte memory and 100 bytes storage
per flop/s are arbitrary values. If someone else
has a better rationale for the hardware requirements
of an AI I'd be interested in hearing about it.
Meanwhile, I'm using them to project when the
runaway acceleration in technology may happen.
Right now I believe that runaway acceleration is
due in this decade, with the Singularity due a few
years after that.
Daniel
__________________________________________________
Do You Yahoo!?
Send FREE Valentine eCards with Yahoo! Greetings!
http://greetings.yahoo.com
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:37 MDT