From: Stephen Reed (reed@cyc.com)
Date: Fri May 31 2002 - 12:10:37 MDT
On Fri, 31 May 2002, Dani Eder wrote:
> I've been tracking the cost of a node in
> a Beowulf-type cluster of commodity computers, where
> the memory is kept at a ratio of 1 byte/flop/sec,
> and storage is 100 bytes/flop/sec.
I likewise have been keeping track of hardware and Moore's law curve. But
for Cyc-style processing I use integer operations, not floating-point. Cyc
runs on a lisp runtime system that we wrote in portable C (gcc). Very few
floating point operations are used; mainly our runtime is pointer chasing,
and type bit-stripping. So I use SpecINT2000 benchmarks which are
available for many current and historical CPUs. Then I set mips =
4*SpecINT2000 and plot mips on a log chart for CPU by year
introduced. Hans Morovac has published his belief that the human brain
runs at 100 million mips.
Regarding disk storage, Cyc-style systems are memory bound and would
operate with modest shared network storage. I would drop storage from the
unit cluster computation, because as you know a linux cluster does not
require disk storage for each node. I would also drop the networking cost
from each unit because you can use a motherboard with an internal
NIC. You can (if you wish) add in a fixed cost for racks, hubs and UPS.
> If optimistic estimates of the required computer
> power for human-level AI are correct at 100 TFlop/s,
> it presently costs $31.9M to buy a human's worth
> of computers.
Here is my calculation assuming AMD Athlon 2000+, boot-up from
network, and omitting fixed costs (not a recommended system, just
low-price parts):
Case: $25 Generic Mid-ATX w/300W power supply
Motherboard + CPU + NIC:
$185 830LR, Athlon 2000+, onboard NIC
DRAM $216 qty 2 PC2700 DDR
total $426/node
$426 / 2896 mips per node = $.147 / mips per node
100 mil mips / 2896 mips per node = 34,530 nodes per 100 mil mips
$426 * 34,530 nodes = $14.7 mil per full time Human Equivalent Computing
> 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 $120K as the total
> cost per year of a technical professional. We are
> therefore about 3.4 doublings in performance/$
> away from economic crossover. At current rates of
> cost reduction, that would be reached in early 2004.
$120K x 5 x 5 = $3 mil crossover.
$14.7 mil is approximately five times your crossover, which by Moore's law
is about four years away (somewhat more than two halvings in cost for
similar performance), about the year 2006.
In the near future the hardware extrapolation is upset by AMD's
introduction of the Opteron AMD32-64 architecture. Kurzweil has argued
that Moore's law will accelerate and this may be evidence. Cyc will use
the 64-bit architecture to contain much larger knowledge bases in RAM. By
this time next year, I expect a hypothetical Opteron single CPU cluster to
achieve Human Equivalent Computing with half the number of nodes,
approximately 17,000. By 2006 (roughly) only 4,000 nodes would be
required, using more powerful CPUs projected to be available. A cube 16
nodes per side would yield 4096 nodes to give you an idea of the physical
space requirement. And because AMD Opterons are designed for N-way
multiprocessor implementations, one might imagine constructing a cluster
in which each node is a 4-way Opteron. Thus 1000 nodes (containing the
4000 CPUs) would occupy a cube only 10 nodes per side.
Note that Human Equivalent Computing is only a comparison between
benchmarked computer performance and theorized human brain computational
power - not any measure of Artificial General Intelligence (AGI).
-Steve
-- =========================================================== Stephen L. Reed phone: 512.342.4036 Cycorp, Suite 100 fax: 512.342.4040 3721 Executive Center Drive email: reed@cyc.com Austin, TX 78731 web: http://www.cyc.com download OpenCyc at http://www.opencyc.org ===========================================================
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:39 MDT