1/16/2001:
SAT-Practice Word
of the Day: assuage
Intermediate Word of
the Day: adjure
Difficult Word of the
Day: blepharitis
Most of O. Henry's short stories are, to my
way of thinking, dated and a bit on the sententious side. But
his humorous short stories (of which the best-known is, "The
Ransom of Red Chief") are the wittiest I've ever seen. One
of his books, "The Gentle Grafters", tells of the exploits
of those two con-men extraordinaire, Jefferson Peters and Andy
Tucker. I enjoyed them so much that in undergraduate school, I
labored at producing my own Chinese copy of an O. Henry "Gentle
Grafters" story, akin to "The Seven Percent Solution"
or Benford's, Bear's, and Brin's sequels to Asimov's 'Foundation"
series. I wasn't very jolly at the time, I was perfectionistic,
and there were no word processors, so everything had to be retyped
every time I made a change. I wasn't able to complete it. But
in the 1980's, I took up the cudgels again, and this time, I finished
the O. Henry knock-off that you're about to read: "The Salting of the Earth".
Sixty
years ago, as a child, I eagerly
anticipated the day when humanity would gain the power to boost
its intelligence and to extend its youth-span. These are both
(I thought) empowerments that will enable us homo-sapients to
advance faster toward our ultimate destiny. It now seems to me
that both of these capabilities are coming into hailing range.
Extending the youth span
Smart Drugs
1/8/2000 Update: My computer failure pushed the "Mind Booster"
review toward the back burner. However, the little spalpeen has
quit kicking and scratching, and tomorrow, I may be able to make
good on my promise.
1/8/2000 Update: Sorry. I spent today's disposable time data-mining
Banner News items. I'll try again tomorrow. (In the meantime,
I've been taking Huperzine-A for a few days, and so far, I've
experienced no ill effects from it (said the window washer, as
he fell past the 40th floor)
Words for the
Techno-Weenie: As stated in
a just-completed computer technology forecast in the latest issue
of Ubiquity
Magazine , IBM is displaying technology that might conceivably
support the present rate of Moore's Law computer technology improvement
through the year 2013*, leading to computers
with 400 times the speeds and 400 times the RAM capacities of
present-day computers (500,000 MIPS, 100,000 megabytes). Today's
Banner News discusses some of the intermediate technological "nodes"
or "insertion points" that we are scheduled to pass
on our way to the computers of 2013. What gives this march of
progress suspense and immediacy is the fact that by 2010, circuit
features would be about 80 atoms wide, by 2020, they would be
only 8 atoms wide, and by 2030, they would be 0.8 atoms wide!
(Ideas like stacking chips in three dimensions don't really help
much because our two-dimensional chips give off so much heat that
they would have to be aggressively liquid-cooled if we were to
try to stack them on top of each other.) Of course, even after
we reach whatever limits of miniaturization are realizable, it's
always possible to increase the degree of parallelism in the processors
themselves, and to employ multiple microprocessors, conceivably
mounting them on the same chip. However, microprocessors typically
generate 20 to 40 watts of power. Eight-to-sixteen such processors
might start to dump too much heat into the computer room.
* - We really only have assurances of continued Moore's-Law
progress through 2005.
An alternative strategy could be to network
processors through fiber-optic cables. One would have access to
remote computational power on demand. Fiber-optic cable capacities
are 1,000-folding every decade, while semiconductors, in accordance
with Moore's Law, are 100-folding every decade. In 20 years, communications
rates could be 1,000,000 times what they are today, with data
swishing back and forth faster than it now sloshes between a microprocessor
and its cache memory. (What happens if a super-intelligent, artificial
intelligence--an Overmind--arises within, or is released into
the network? Whoo-oo-oo-oo-oo! Happy Halloween!)
One Banner News article deals with the fact
that as the number of atoms in a transistor declines, the point
is reached where there aren't enough silicon atoms for the few-parts-per-million
trace elements (dopants) that convert the silicon into a semiconductor.
Once we have only 250,000 silicon atoms in our transistor, there
will only be one trace atom present. And if we get below that
threshold.... Carbon nanotube transistors have been suggested
as a far-out alternative.
The theoretically-smallest carbon nanotubes are
about 0.4 nanometers (4 Angstroms) in diameter. That's pretty
small. That would correspond to the design features we would require
in 2025 if we were to stay on a Moore's-Law shrinkage curve. Of
course, right now, such a number has no practical significance
except to suggest that, perhaps, carbon nanotubes needn't be ruled
out as very tiny circuit elements.
It also seems to me that the days of conventional
silicon circuits must surely be numbered. We have shrunk
transistor dimensions five orders of magnitude in five decades
(from 18 millimeters in 1950 to 0.18 micrometers in 2000). Three
more orders of magnitude would take us slightly below atomic dimensions.
Conventional silicon depends for its fortuitously-convenient properties,
upon bulk-silicon effects. Although other materials might be mounted
on silicon substrates, bulk-silicon effects must surely be reaching
their miniaturization limits. Other technologies may take us down
to the atomic level, but it won't (I think) be done with bulk-silicon.
Another vital factor in these deliberations is that
of manufacturing costs.
It's interesting to compare some of these sizes
with the sizes of organic entities. Cells are of the order of
10-to-20 microns in length. Our memory cells are currently of
the order of 0.4 microns in length with 0.18-micron circuit features.
We are approaching production of 1,000,000,000-transistor memory
chips, using memory cells 0.05 microns across. Small organic molecules
are of the order of 0.001 microns-or-less across, or less than
1/10,000th the length of a cell. The sizes of the simple molecules
that comprise a cell stand in relation to the size of a cell the
way that our individual cells stand in relation to us. They're
about 1/10,000th the size of a cell, just as our cells are about
1/10,000th our size. A cell is comprised of about 1,000,000,000,000
molecules.
The human brain harbors 100,000,000,000
neurons. By 2010, if we remain on a Moore's Law curve, we should
be able to place 100,000,000,000 transistors on a chip. One artificial-neuron
design uses 7 transistors to emulate a neuron. At that rate we
would need 7 chips to emulate the brain. However, the same artificial
neuron design that uses 7 transistors to emulate a neuron requires
5 transistors to replace a synapse, and there are, perhaps, 10,000
synapses for every neuron. (5,000 synaptic pairs) So it would
take fifty thousand,100,000,000,000-transistor chips to emulate
the brain. On the other hand, computer chips are a billion times
faster than neurons, and probably a million times faster than
synapses. Also, transistors are relatively reliable devices compared
to neurons and synapses. If we're willing to slow our electronic
neurons and synapses to 1/50,000th of their top speeds, we might
be able to place 50,000 of them together in a small package without
generating too much heat to permit such a concentration.
The reason this is so important is because rapidly-rising computing power forms one of the cornerstones of our economy. It will also form the cornerstone for robotics and artificial intelligence, so limitations upon its continued improvement are very important in casting humanity's technological horoscope. In evaluating what's happening, it's important to note that not only are costs, sizes, and power requirements dropping exponentially, computer speed can rise exponentially because of the exponential reductions in size and power requirements. However, exponential growth can't continue through terribly many ten-foldings before hitting some wall. The table below shows some of the characteristics of future chips. The second column, ATOMS, gives the sizes of features--e. g., the widths of transistors--measured in atomic diameters. Column 3, ATOMS/BIT, shows very roughly how many atoms would be available per 1-bit memory cell. This includes conductors to access the memory cells and insulators around the electrically-charged elements of the cells. At 8,000,000 atoms, there would still be enough atoms to dope the silicon with trace elements to render it semi-conducting. BITS/CHIP specifies the number of bits that can be stored on a state-of-the-art RAM chip at each point in time (109 bits today). SPEED, GHz is a very simplistic set of extrapolations that assume that smaller dimensions would support the blistering clock speeds tabulated in Column 5. (They probably won't.) Column 6, TRANS/CHIP, is a fanciful tabulation of the number of transistors that might be mounted on future microprocessor chips.
| YEAR | ATOMS | ATOMS/BIT | BITS/CHIP | SPEED, GHz | TRANS/CHIP |
| 2000 | 800 | 8,000,000,000 | 109 | 1 | 42,000,000 |
| 2010 | 80 | 8,000,000 | 1011 | 100 | 1,000,000,000 |
| 2020 | 8 | 8,000 | 1013 | 10,000 | 25,000,000,000 |
| 2025 | 3 | 320 | 1014 | 100,000 | 125,000,000,000 |
| 2030 | 1 | 8 | 1015 | 1,000,000 | 500,000,000,000 |
I don't expect
the numbers in this table to be realizable. I merely present them
to show what would be required if computer technology were to
continue in the same channel in which it has traditionally run
over the last 50 years.
Years ago, 300 GHz was the highest frequencies
at which electronics devices had been made to run. 10,000 GHz
goes with wavelengths in the far infrared, corresponding to a
color temperature of the order of 100º K. 100,000 GHz is
the middle of the infrared spectrum, corresponding to a wavelength
of 3 m. 1,000,000
GHz is in the near-ultraviolet, representing a wavelength of 3,000
Angstroms.
UNIVAC 1 used a mercury delay-line memory that
stored 1,024 36-bit numbers that passed through the computer about
1,000 times a second. Could one future possibility be a fiber-optic
delay line that would transmit 1012 - 1015 bits through the computer 1,000,000,00 to 30,000,000,000
times a second? That would be the modern equivalent of the old
mercury delay line. We should soon have digital circuitry that
could accommodate those speeds.
Intel is on record hoping that new approaches
to microprocessor technology will come to the fore after 2010.
I could imagine other memory concepts, such as optical storage
of one kind or another, having both the speed and capacity to
possibly continue when semiconductor memories reach their elastic
limits.
To get across the mind-expanding changes that
would be required to follow a Moore's Law curve through the year
2030, if we imagine our Year-2000 transistors (which are already
too small to be seen by the most powerful optical microscopes)
to be one meter across, then we will have to shrink our transistors
to one millimeter (1/25th-inch) across by 2030. One million 2030
transistors would have to fit in the invisibly-small space currently
occupied by one transistor. Even to remain on a Moore's-Law curve
through 2010, we'll have to shrink our dimensions until we can
pack 100 transistors into the space occupied by one transistor
today. Most of the discussions about this that I'm seeing say
that extreme-ultraviolet lithography should carry us through the
rest of the lifetime of silicon semiconductor technology "as
we know it".
One bright spot on this horizon is IBM's announcement
that they've found a way to reduce transistor surface widths to
0.01 m. That
would carry us through a Moore's Law progression to 2013. And,
as I've mentioned, there should be at least 5 more years of improvements
after that, as prices fall and multiprocessor systems become cheaper
and more popular.
Another way to look at the challenges
that lie ahead is in terms of heat and power dissipation. Suppose
that in ten years, we hundred-fold the number of transistors we
can mount upon a single chip, by making them 1/10th as wide as
they are today (requiring 1/100th the chip real estate). If we
assume that the energy required to switch a transistor is proportional
to its area, then each transistor will dissipate only 1/100th
as much energy when it switches states as today's transistors.
But since there are 100 times as many of them on chip, the total
energy dissipated by the chip during each clock cycle
will be the same as it is today. But we hope to run them 100 time
as fast as today's chips, whizzing along at a 100 GHz clock speed.
That means that they would require 100 times as much power as
today's chips, and would dissipate 100 times as much heat. In
reality, transistors probably shrink in depth as well as length
and width, so, since their volumes may be 1/1,000th what they
are today, they might only dissipate 10 times the power. Of course,
that's still not permissible. Today's fastest microprocessors
have to be aggressively cooled. One of the ploys used to lower
heat dissipation is to run the chips at lower voltages. But now,
there's another barrier. Random thermal energy at room temperature
is about 1/40th of an electron-volt. If a transistor voltage is
to act as a barrier to electrons, it must be many times that 1/40th
volt, or too many electrons will climb over the voltage "hill".
Also, the change in energy involved in a transistor energy transition
must be many times the 1/40th-of-an-electron-volt background noise.
Now consider a chip that has 10,000,000,000 transistors and is
running at 10 GHz. That's about where we expect to be in four
more years. Suppose we say that the maximum amount of power (and
heat) that it can be allowed to dissipate into our computer room
is 10 watts. If we equate 10 watts to 10 billion transistors switching
10 billion times a second, the amount of energy in an energy transition
in each transistor each time it switches must be 10/1020
joules or about 10-19 joules. But 1/40th of an electron-volt
is about 1/25th X 10-19 joules, or about 1/25th of
the maximum switching energy that's compatible with 10 watts of
power dissipation if that many transistors are switching
10,000,000,000 times a second. Of course, in a microprocessor,
not every transistor is going to switch on every cycle.
If we were to ramp this up to the 2010 era
and were to envision 100,000,000,000 transistors running at 100
GHz, we'd be faced with a "speed-power" product that's
100 times as great as it will be in 2005, or 10,000 times greater
than it is now. To continue in this vein to 2020, the number of
state transitions would 100,000,000-fold. And to make it to the
projected 2030 level, this number would rise by a factor
of 1,000,000,000,000! Of course, in going from a 3.6 megahertz
clock speed, with a transistor count of 235,000 in the latter
80's, to a 1,500 megahertz clock speed, with a transistor count
of 42,000,000, we've already increased transistor switching per
chip per second by a factor of about 75,000. On the other hand,
those switching energies weren't remotely close to 1/40th of an
electron-volt.
It's going to be a challenge.
We're still far above speed-power products
that are subject to the Heisenberg Uncertainty Principle
Mike
Hess Comments Anent the "Severely Gifted"
Patrick
Wahl's Discussion of the Word "Nonce"
Patrick's
Question About Plans for the Website
AFTER THE FIRE
Upon this hill laid bare to sun
The bee attends the flowers,
And by the toppled chimney stone
The shadow tells the hours.Domestic as the kitchen clock
That measured out our day
Before the foe bypassed the lock
And found a secret way.The spider who had claimed the wall
Now claims the sapling pine,
And angles for his evening meal
With silver reel and line.The mouse finds other store of grain,
The sparrow other eaves,
The cricket shakes his tambourine
Beneath the maple leaves,As heedless of his brother's shout
As we were heedless then
Before the enemy without
Became the foe within.--"Window to the South."
Vivian Smallwood