Re: Hardware Progess: $165/Gflop

From: Eugen Leitl (
Date: Fri Jan 02 2004 - 12:45:08 MST

On Fri, Jan 02, 2004 at 08:37:35AM -0800, Dani Eder wrote:
> The cost of processing power continues to drop at
> a Moore's Law pace. Cost has dropped 22%

I hope the rest of your post is not based on this premise, because it's flat
out wrong. Moore is about integration density, not performance. Performance,
unfortunately, does not linearly fit semilog plots, and hasn't been for quite
a while now.

> 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

How do you know which sort of system is useful for AI work?
I happen to think you'd need a low latency packet switched signalling fabric,
and an CPU/FPGA core with very-large-bandwidth embedded memory.

Can you tell me where I can currently buy such system? No?

> NIC: 14 4x10/100 Ethernet cards
> Wiring: 8 4x6ft Cat 5 cables w/RJ45 connectors
> Hub: 10 1/6x24 port fast ethernet switch

Ethernet is lousy latence-wise (30 us best case), but I wonder why you didn't
include GBit Ethernet in your comparison.

> AMD processors when used in clusters reported in
> the supercomputer list achieve about
> 1.375 Flop/Hz out of a theoretical 2 Flop/Hz.

You can as well cite random numbers from the back of a cereal box.
While it is obvious, that AI is a large-scale numerics application, it is not at all
obvious which kind it is.

> Note that many announced supercomputers
> have higher costs per Gflop due to installation,
> software, and facility costs. This is a best case

Hardware costs do not feature prominently in large scale commodity clusters:
here power, air conditioning and administration dominate. You can partly
forego air conditioning by loose shelving in a hall with fans in the roof,
but the rest of the costs are not that easy to reduce.

> 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

This assumes that 17 M$ worth of hardware achieves equivalent real-world
performance as a human. In other words, it assumes AI is a solved problem.

Now this is a bit premature.

> 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

Please don't use garbage OPs. They're tied to a specific benchmark, and that
gives you another two zeroes to add to your estimate.

Computers are not just CPUs. These are usually memory-starved, and here
things like memory bandwidth, latency, network latency and crossection
bandwidth come into play.

> The trend predicts highly automated production should
> arrive around 2020. At some point around then, it
> should be possible to design replicating industrial

I notice that no one is currently attempting to build self-rep systems,
despite seminal work being about half a century past. It's useless to
speculate when something can be done, if the humanity collectively decides to
not pursue that particular path.

> 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
> year).
> Therefore even if improvements in computer components
> slows down due to limits of semiconductor scaling,

You're completely ignoring molecular circuitry, and scaling issues of very
large systems.

> their performance/cost ratio may continue to improve
> due to highly automated production.

-- Eugen* Leitl leitl
ICBM: 48.07078, 11.61144
8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE

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