Estimates for hardware requirements?

From: Dani Eder (
Date: Wed Nov 14 2001 - 10:04:51 MST

IBM has announced:

the first computer to be built over the threshhold
of Moravec's estimate of human-equivalent processing
power (200 TFlops for Blue Gene/L vs. Moravec's
figure of 100 TFlops).

In light of that, is it time to start thinking about
the rest of the system required to develop the seed
The software is necessary, but not sufficient for
a complete system.

I can see at least three levels of computer set-ups
that might be required:

(1) Workstation(s) to do code development on
(2) Test system to test parts of the seed AI code
    or early stages of it's run
(3) Full system capable of getting to

For the first level, a specially designed system
probably isn't required. But for the second &
third, a complete system will require thinking

- Core computer hardware requirements (Flops,
RAM, disk capacity, internal and external data
transfer rates)

- Interfaces or barriers to the outside world

- Facility support (floor space, UPS, electrical
and HVAC to support your 1000 linux box cluster
or whatever)

- Staff

- Procedures (do you do daily backups? what do
you do if the console operator starts acting like
a zombie? etc.)

In any engineering project, the first question is
what are the requirements for the project. So my
question for the list is has anyone got a better
estimate on the hardware requirements than the ones
based on human brain capacity (which range from
100 - 100,000 TFlops depending on who you ask)?

If you have multiple approaches to AI, even a
rough estimate of the requirements can help guide
what you do now. For example, say the CYC-like
approach of feeding in millions of bits of
knowledge takes 1 TFlop to run, and a neuron
simulation takes 1000 TFlop to run. Given
limited financial resources, you might not even
have to start development of the latter for
5 years, since the hardware won't be affordable
for 10 years. That means you can spend the next
5 years on the former approach on the chance it
might work, rather than just waiting for the
hardware to get cheap enough for the latter.

That's a simplistic example, but I hope it gets
the point across that there is value in even a
rough estimate of the requirements of an AI.


Do You Yahoo!?
Find the one for you at Yahoo! Personals

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:37 MDT