From: Brian Atkins (brian@posthuman.com)
Date: Fri Jun 22 2001 - 13:21:56 MDT
Dani Eder wrote:
>
> The latest list of the top 500 supercomputers in the
> world has just been published (www.top500.org). The
> sum of all 500 machines now is at 108 Tflops, which,
> using Moravec's estimate, is equivalent to 1.08 human
> brain power. This is a 23% increase from the previous
> list published 6 months ago.
>
> There is a nice graph of the trends at:
> www.top500.org/lists/2000/11/slides/ListOverTime.pdf
>
> Note that the time for the N=500 machine to reach the
> previous power of all top 500 machines is about 12
> years. Using N=500 as a proxy for a machine that
> an AI project can be expected to get it's hands on,
> we're looking at 2013 for a 'possible singularity
> date'. Another 7 years brings the N=500 machine to
> the level of an average desktop PC. So if asked for
> the expected date of the singularity, I would now
> give a range of:
>
> Earliest: Now (ganging up many top supercomputers
> or SETI@home type distributed system)
>
> Early: 2013 (Research AI project gets it's hands
> on 100 Tflop machine)
>
> Late: 2020 (If it takes a full neural simulation
> level of computational power = 3000 Tflop)
>
> Latest: 2027 (Full neural simulation on a PC, and
> no acceleration in Moore's law until then)
>
> On the related issue of how much will a machine cost,
Thanks for the update Dani. Personally I feel that your dates should
be moved up a bit. For instance, we expect Blue Gene-level hardware to
be available circa 2005. The development cost of this project is said
to be $100m, but if you wanted to buy one for yourself it would be less.
And if you are saying we only need 1/10th of its power for Moravec's
estimate, then the cost could be in the $2 to 5m range? Then consider
that we could develop a seed prototype on a smaller system, and then
rent some time on a supercomputer to run the final tests. So the actual
costs for this might drop below $1m in the 2005-2010 timeframe.
-- Brian Atkins Director, Singularity Institute for Artificial Intelligence http://www.intelligence.org/
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:36 MDT