Re: Hardware Cost

From: Brian Atkins (
Date: Wed Apr 25 2001 - 11:44:44 MDT

Dani Eder wrote:
> I've been keeping track of the component costs,
> and a current version can produce 122 GFLOPS
> for $32,400 in parts, or $266/GFLOPS.
> I'm curious how this type of Linux cluster compares
> to the processor requirements for a Web-Brain
> implementation or an estimated size for a seed AI
> (if anyone has taken a stab at that yet).
> Another factor of 10 reduction in cost would put
> a one-human neuron-model type machine (3000 TFlops)
> within reach of the amounts being spent on the
> largest government supercomputers.

The planned IBM Blue Gene supercomputer has a projected cost of $100,000,000
for 1,000,000 gigaflops, or $100/gflop. If they build more of them after
that for other people the cost will likely be much less since all the design
work would have already been done. So don't worry, your needed factor of 10
cost reduction should be accomplished around 2006.

Unfortunately there doesn't seem to yet be a concrete idea of how many
ops/sec are required for a decent seed AI, or how much of an impact the
required AI code will have on those ops. If it turns out that the only
way to get a working seed is via some certain kind of code that totally
bottlenecks on the bus/memory speed, then those theoretical 1000 peak
teraops may not mean much.

SIAI doesn't even have any code written yet; in a year or two we will probably
have a better idea of what we might need. Perhaps Ben could give a guesstimate
of what he thinks "Webmind 2"'s requirements might be to achieve some decent
level of general intelligent behavior?

Brian Atkins
Director, Singularity Institute for Artificial Intelligence

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:36 MDT