AI hardware was 'Singularity Realism'

From: Keith Henson (
Date: Sat Mar 06 2004 - 09:54:03 MST

At 10:59 PM 05/03/04 -0500, Beb wrote:


>Of course, if I speak from my gut, I must admit don't even think it
>would take a Manhattan Project style effort. I think I could get there
>in 2-3 years with an investment of $20 million for staff and hardware.

If you can make a strong case I think it could be funded without much of a

But I don't think anyone right now has a clear idea of what it would take
to build an AI by other means than a simulation of a human brain. (Correct
me and send pointers if someone *does* know.)

If I were to lay out a research project leading in the simulation
direction, I would probably try to build on William Calvin's models of what
is going on in the human cortex. A few years ago I worked out how large an
area of computers it would take to simulate the cortex, but I can't find
anything so perhaps I didn't put it on the net.

The human cortex surface area is about 1/4 square meters (half a meter on a
side), packed solid in a hexagonal array with "active elements," groups of
about 100 neurons that act together in a 0.03 mm minicolumn. (See page 29
in _The Cerebral Code_, webbed here:

I think with some simplification the active elements (and their
connections) could be simulated at a millisecond update rate with a
specialized 1 square cm processor running at about a GHz and dissipating
about ten watts. (Current x86 processors are in this size and power range
for this clock rate.)

The linear expansion factor from 0.03 mm to 1 cm is about 300, so this
would amount to a silicon surface of about 150 meters on a side--which is
close enough to my memory of 200 meters on a side last time I worked it
out. There would be 10,000 processors per square meter, or 225,000,000 of
them in total.

Power would be interesting, 2.25 Gw--which is about 1/3 of the peak power
output of Grand Coulee Dam. At 2 cents a kwh, the power bill would be $50k
an hour, or $400 million a year. Cooling would be interesting
too. Cost? If the processors were 1/10 of the total, and could be
produced for $10 each (perhaps--processing silicon takes close to a dollar
a square cm just for electricity) then such an AI project would run about
$22 billion for hardware--which is only a thousand times off from Ben's

Makes a human brain at 25 watts and fitting into a 1500cc skull *really*
impressive. It's on the order of 100 million more efficient in power and
volume then current solid state hardware--which is fairly impressive itself!

It would take some thinking about this kind of project, but the hardware
organization might keep it from a "takeoff." Of course that makes it less
useful as well as less dangerous.

Keith Henson

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:46 MDT