Hardware Progess: $165/Gflop

From: Dani Eder (danielravennest@yahoo.com)
Date: Fri Jan 02 2004 - 09:37:35 MST

The cost of processing power continues to drop at
a Moore's Law pace. Cost has dropped 22%
in the past eight months and by one half
in the past 20 months.
I've been tracking the cost of a node in
a "Sparse Flat Network Neighborhood" type
cluster of commodity computers. The basis
machine is described at:

Their machine does not have much memory and
no storage, so for the purpose of tracking the
cost of a system useful for AI work, I track
nodes where the memory is kept at a ratio of
near 1 byte/Hz, and storage is 100 bytes/Hz. The best
cost/performance nodes I have found are as follows:
Date $/Gflop/s
Nov 01 600
Jan 02 437
Mar 02 397
May 02 319
Aug 02 272
Oct 02 244
May 03 212
Jan 04 165
Component prices are obtained from pricewatch.com,
including shipping but not sales tax. Current
configuration is as follows:
Case: $20 Generic Mid-ATX w/300W power supply
Motherboard: 38 MSI K7T Turbo 2
CPU: 59 AMD Athlon XP2200 (1.8 Ghz)
Memory: 128 3x512MB PC133 SDRAM
Storage: 131 180 GB EIDE internal Hard Disk
NIC: 14 4x10/100 Ethernet cards
Wiring: 8 4x6ft Cat 5 cables w/RJ45 connectors
Hub: 10 1/6x24 port fast ethernet switch

Total $408/node

AMD processors when used in clusters reported in
the top500.org supercomputer list achieve about
1.375 Flop/Hz out of a theoretical 2 Flop/Hz.
Therefore the performance/$ is:

$408 / (1.8 GHz * 1.375 Flop/Hz) = $165/Gflop/s

Note that many announced supercomputers
have higher costs per Gflop due to installation,
software, and facility costs. This is a best case
lower bound for a self-assembled system.

If optimistic estimates of the required computer
power for human-level AI are correct at 100 TFlop/s,
it presently costs $16.5M to buy a human's worth
of computers. I have estimated an 'economic
crossover' of $3M for 100 TFlop/s when computers
become generally cheaper than humans. This is
based on a computer being able to put in 5x as
many productive hours as an average human, a 5 year
payback time on the hardware, and $120K as the total
cost per year of a technical professional. We are
therefore about 2.5 doublings in performance/$
away from economic crossover. At the recent trend
of cost reduction, 3.3% per month, that would be
reached in 52 months, or May 2008.

Economic crossover does not imply we would have
human-level AI at that point, but rather that
skilled human tasks which require lots of processing
power would be cost-feasible. In particular, high
level factory automation should be possible. This
has implications in the social and economic realms.

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. Assume the most powerful
machine available for AI work is in the $3 M price
range. Then the expected date of a human-level
machine is 2008-2024. The expected date for
superintelligent machines is 7.5 years later. This
is calculated from a human taking about 20 years
to train. A machine 8 times faster than a human
is expected 5 years after a human-equivalent, and
it would be expected to train in 20/8 = 2.5 years.

Once superintelligent machines are available, the
time constant for doubling performance may drop to
much shorter periods as superintelligent machines
design their successors ever more rapidly. This
singularity may occur in the 2015-2033 time period.

Labor productivity trends also point to a singularity
in the ~2020 time period. From 1992 to 2003, labor
input in durable goods manufacturing has fallen by 42%

per unit of output. Durable goods are defined as
having a useful life of over a year, as opposed to,
say food, which gets consumed quickly. In particular
factory equipment, such as machine tools, that are
used to produce durable goods, are themselves durable

The trend predicts highly automated production should
arrive around 2020. At some point around then, it
should be possible to design replicating industrial
systems, either single factories or networks of
factories, that can produce copies of themselves
as well as useful products with little human input.
Preliminary estimates indicate that these replicating
factories may have a time to copy themselves on the
order of 1 year, which is similar to the rate of
improvement of computer power, and much faster than
the rates of economic growth we are used to (0-10%per

Therefore even if improvements in computer components
slows down due to limits of semiconductor scaling,
their performance/cost ratio may continue to improve
due to highly automated production.


Do you Yahoo!?
Free Pop-Up Blocker - Get it now

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:43 MDT