AN EXCURSION:
THE FUTURE DESIGN OF MACHINES THAT THINK
In the previous chapters we have described four giant mechanical
brains finished by the end of 1946: Massachusetts Institute of
Technology’s Differential Analyzer No. 2, Harvard’s IBM Automatic
Sequence-Controlled Calculator, Moore School of Electrical
Engineering’s Electronic Numerical Integrator and Calculator (Eniac),
and Bell Telephone Laboratories’ General-Purpose Relay Computer. All
these brains have actually worked long enough to have demonstrated
thoroughly some facts of great importance.
WHAT EXISTING MACHINES HAVE PROVED
The existing mechanical brains have proved that information can be
automatically transferred between any two registers of a machine.
No human being is needed to pick up a physical piece of information
produced in one part of the machine, personally move it to another part
of the machine, and there put it in again. We can think of a mechanical
brain as something like a battery of desk calculators or punch-card
machines all cabled together and communicating automatically.
The existing mechanical brains have also proved that flexible,
automatic control over long sequences of operations is possible. We can
lay out the whole routine to solve a problem, translate it into machine
language, and put it into the machine. Then we press the “start”
button; the machine starts whirring and prints out the answers as it
obtains them. Mechanical brains have removed the limits on complexity
of routine: the machine can carry out a complicated routine as easily
as a simple one.
The existing giant brains have shown that a machine with hundreds of
thousands of parts will work successfully. It will operate accurately,
it will run unattended, and it will have remarkably few mechanical
troubles.
These machines have shown that enormous speeds can be realized: 5000
additions a second is Eniac’s record. High speed is needed for many
problems in science, government, and business. In fact, there are
economic and statistical problems, now settled by armchair methods,
for which high-speed mechanical brains may make it possible to compute
answers rather than guess them.
Also, these machines have been shown to be reasonable in cost. The cost
of each of the large calculators is in the neighborhood of $250,000 to
$500,000. If we assume a ten-year life, which is conservative, the cost
is about $3 to $6 an hour for 24-hour operation. Since each mechanical
brain can, for problems for which it is suited, do the work of a
hundred human computers, such a machine can save its cost half a dozen
times. And these machines are only engineers’ models, built without the
advantages of production-line assembly.
The cost of giant mechanical brains under design in 1947 and 1948
is in the neighborhood of $100,000 to $200,000. The main reason for
the reduction from the previous cost is the use of cheaper automatic
memory. As designs improve and charges for research and development are
paid off, the cost should continue to go down.
NEW DEVICES FOR HANDLING INFORMATION
In the laboratories working on new mechanical and electronic brains,
scientists are doing a lot of thinking about new devices for handling
information. Research into devices for storing information shows that
_magnetic wire_ as used in sound recording is a rather good storage
medium.
Magnetic Wire
For example, on a hundredth of an inch of fine steel wire we
can “write” a _magnetized spot_ by means of a small “writing”
_electromagnet_. The electromagnet is simply some copper wire coiled
around some soft iron shaped in a U. When current flows through the
coil, the iron becomes a magnet, and the tips of the U magnetize the
little section of the wire between them. The magnetized spot can be of
two kinds, say north-south or south-north, depending on which way the
current flows. We can “read” this difference by means of another small
“reading” electromagnet. We can erase the spot by means of a stronger
“erasing” magnet that produces a uniform magnetic state throughout the
wire. The difference between north-south and south-north corresponds
to the difference between 1 and 0, or “yes” and “no,” etc., and is
a _unit of information_ (see Chapter 2). Many other variations are
possible. For example, the presence or absence of a magnetized spot may
be the unit of information, or the “writing,” “reading,” and “erasing”
electromagnets all may be the same.
Magnetic wire sound recordings made in the 1890’s are still good.
This fact shows that magnetic wire may be a more permanent medium for
storing information than is paper. Stray magnetic forces are likely
to have no harmful effect on information stored on magnetic wire, for
these forces would not be strong enough or detailed enough to change
greatly the difference between the magnetized spot and its neighboring
neutral area.
A reel of magnetic wire a mile long and ³/₁₀₀₀ of an inch thick costs
about $5. At 80 magnetized spots to the inch, a mile of wire can store
about 5 million units of information. Hence, the cost of storing one
unit of information is about ¹/₁₀₀₀₀ of a cent. The time needed for
changing a magnetized spot from 1 to 0 or from 0 to 1 is about ¹/₁₀₀₀₀
of a second.
Magnetic Tape
There is, however, a storage device that may be even more useful, and
this is _magnetic tape_ (see Fig. 1). The usual size of such tape is ¼
inch wide and 2 or 3 thousandths of an inch thick. Magnetic tape may be
made of plastic with magnetic powder all through it, or it may be of
paper coated with magnetic powder, or it may be of stainless steel or
a magnetic alloy, or it may be of brass or a nonmagnetic alloy coated
with a magnetic plating.
Magnetic tape has the added advantage that from 4 to 20 channels across
the tape can be filled with magnetized spots, and the cost then becomes
about ¹/₁₀₀₀₀₀ of a cent per spot. It seems possible that 1000 units
of information can be stored in a quarter of a square inch of magnetic
tape. This means that more than 1 million units of information can be
stored in a cubic inch of space filled with magnetic tape, and about 2
billion units of information in a cubic foot, except that some of the
space should be allotted to the reels and other equipment that hold
the tape (see Fig. 2). This is closer packing than printed information
in the telephone book, and yet with magnetic tape we can get to the
information automatically.
[Illustration: FIG. 1. Magnetic tape.]
[Illustration: FIG. 2. Tape reels.]
Think of the enormous files in libraries, government, and business.
Think of the problems of space and cost and access which these files
imply. We can then see that this new development may well be of
extraordinary importance.
Mercury Tanks
[Illustration: FIG. 3. Mercury tank.]
Scientists are investigating other storage devices having still more
remarkable properties, but these have the disadvantage that, when the
power goes off, the information vanishes. One of these new storage
devices is called a _mercury tank_ (see Fig. 3). It consists mainly
of a section of iron or steel pipe filled with mercury. At each end
of this pipe, touching the mercury, is a thin slab of a crystal of
_quartz_. Quartz, which is a common stone, and which nearly all sand
is made of, changes its shape when pulsed with electricity. We put a
pattern of electrical pulses into the quartz slab at one end of the
mercury tank; for example, we could have the pattern 1101 meaning
“pulse, pulse, no pulse, pulse.” The electrical pulses going into the
quartz slab make the quartz vibrate. Thus ripples are produced in the
mercury, and waves in the pattern 1101 meaning “wave, wave, no wave,
wave” travel down the tank and strike the quartz slab at the far end.
The quartz slab there changes its shape in the rhythm 1101, and it
converts the waves back into electrical pulses in the same pattern.
Then we take the pulses out of the far end along a wire, make them
stronger again with an amplifier, give them the right form again, and
feed them back into the front end of the mercury tank. The mercury tank
is a clever use of the principle of an _echo_, as when you call across
a valley and the rocks answer you back. We can store a pattern of 400
pulses (each a unit of information, a 1 or a 0, and each a millionth
of a second in duration), in a mercury tank about 20 inches long. A
mercury tank and an echo are examples of _delay lines_—“lines” along
which waves are “delayed.”
Electrostatic Storage Tube
Another of the memory devices being developed is called an
_electrostatic storage tube_ (see Fig. 4). This is a big electronic
tube with a _screen_ across one end. The screen may be of two layers:
one of copper, which conducts electricity, and one of _mica_, a
material that does not. In the other end of the tube is a _beam_ of
electrons, which we can turn on and off and shoot at any of 2 or 3
thousand specific points or _spots_ on the screen.
[Illustration: FIG. 4. Electrostatic storage tube.]
There are two sizes of _electric charge_ or quantity of electrons; we
can call these 1 and 0. In about a millionth of a second, we can put
either size of charge on one of the spots of the screen. With other
circuits we can keep it there as long as we want, if the power does
not flicker off. We can “remember” perhaps 2 or 3 thousand units of
information in one of these electronic tubes. We can read, write, or
erase any unit of information in a few millionths of a second.
Neither the mercury tank nor the electrostatic storage tube had, by the
end of 1947, been put into a working mechanical brain. But there is
good reason to believe that they will be successful devices and will
open up a new era of speed in storing and referring to information.
In fact, several laboratories are developing electronic calculating
circuits using these devices which will perform up to 100,000 additions
a second or 10,000 multiplications a second. Our minds certainly
stagger at the thought of such speeds.
NEW OPERATIONS
Many kinds of combining operations have already been built into one or
more mechanical brains. The operations may be arithmetical: addition,
subtraction, multiplication, division, looking up numbers in tables,
etc. Or the operations may be logical: comparing, selecting, checking,
etc. Additional logical operations will be built into some of the
mechanical brains now being constructed: sorting, collating, matching,
merging, etc.
NEW IDEAS IN PROGRAMMING
_Programming_—the way to give instructions to machines—is also being
studied in the laboratories. Several new ideas of importance have
developed as a result.
One idea is that the machine should be able to store its instructions
or _program_ or _routine_ in its memory in just the same physical ways
as it stores numbers. There is basically no reason why numbers only
should be stored in some registers, and instructions only stored in
other registers.
Another idea is that the machine should have in its permanent memory
any subroutine it may need. For example, a subroutine should always be
available in the machine for finding _square root_. At any time when a
square root was needed, we would only have to call on the machine for
the subroutine of square root. The machine would then consult the right
part of its memory and carry out the subroutine for square root.
A third idea, and one of the most interesting, is that the machine
should be able to compute its own instructions. For example, consider a
program for finding the product of two _matrices_ (see Supplement 2),
each of 100 terms in an array of 10 columns and 10 rows, resulting in a
new _matrix_ of 100 terms. The whole program can be made to consist of
about 50 orders. Only one of them is “multiply,” and only one of them
is “add”; the other orders consist of how to choose expressions to be
multiplied or added, etc.
Such problems as these are often fascinating to mathematicians, who
love to play with the intricate ideas needed.
NEW IDEAS IN RELIABILITY
Reliability has a number of aspects:
1. No wrong results allowed out of the machine.
2. Few failures.
3. Rapid location of failures.
4. Quick repair or replacement of parts that fail.
5. Easy maintenance.
6. Unattended operation overnight.
For example, Bell Laboratories proved that mechanical brains can be
built so that no wrong results are allowed to come out. In other words,
the machine checks itself all the time as it goes along and stops at
once if the check shows that something is wrong. This is likely to be a
standard feature of new automatic thinking machinery.
The frequency of failures in the machinery being designed in the
laboratories may be of the order of one or two mechanical failures
a week. For any type of failure an alarm circuit and trouble lights
will show what part of the machine needs attention. Plug-in parts for
replacement are already in use in at least two of the four mechanical
brains described and should be available in all the new machines. It is
possible to build a machine that will automatically change from failing
equipment to properly functioning equipment. For some years though,
this may be too expensive to be reasonable.
The use of magnetic tape for storage reduces greatly the number of
parts and so increases reliability. For example, instead of 18,000
electronic tubes in an electronic brain, there may be less than 3000.
A final degree of reliability is gained when most of the time the
machine operates unattended. Then, there is no human operator standing
by who may fail to do the correct thing at the moment when the machine
needs some attention. In fact, the motto for the room housing a
mechanical brain should become, “Don’t think; let the machine do it
for you.” Unattended operation from the end of one working day to
the beginning of the next, with the machine changing itself from one
problem to another problem, has already been proved possible on the
Bell Laboratories machine.
AUXILIARY DEVICES
In order to use a mechanical brain, we have to give it and take from it
language that it understands, _machine language_. A mechanical brain
that can do 10,000 additions a second can very easily finish almost
all its work at once. How can we, slow as we are, keep our friend,
the giant brain, busy? We have found so far several answers to this
question, none of them yet very good.
Devices for preparing input will be very important. For each brain, we
shall need a great many of these devices. For, at best, we type at a
rate, say, of 4 characters a second, selecting any one of some 38 keys,
each of which is equivalent to about 6 units of information. This is
about 800 units of information per second. The machine, however, is
likely to be able to gulp information from its input mechanism at the
amazing rate of 60,000 units of information per second, equal to 75
people typing with no mistakes and no resting. Fortunately, at least
some of the time the machine will be busy computing!
For an input-preparation device, we may get something that can be
fastened to an ordinary typewriter and that will produce magnetic
tape agreeing with what is printed by the typewriter. Since the input
information must be carefully verified, we shall need a second magnetic
tape device such as exists for paper tape on the Bell Laboratories
machine: the _processor_. The processor takes two hand-prepared tapes,
compares them, reports any differences, and produces a third tape. The
third tape copies the two original tapes if they agree, and it receives
corrected information as furnished by a girl at a keyboard if the two
original tapes disagree.
For information already on punch cards, we need an input device that
will read punch cards and write on magnetic tape. Where information is
on punched paper tape, we need a machine that will read punched paper
tape and write on magnetic tape.
Problem data, tables of numbers, and routine instructions will go
into the mechanical brain. They will all be prepared on regular input
devices. The machine will accept information in the form in which it is
most convenient for you and me to prepare it. Then, the machine will
be instructed to change the information into the form with which it is
most convenient for the machine to operate.
Many output devices will also be needed, since the machine will be able
to produce information very swiftly. These output devices might be
cabled to the machine. A kind of traffic control system would govern
them. Each will have a magnetic tape that will be loaded up swiftly
with information. Then the output device will unload its information
more slowly, in any form that we may desire: printing, graphs, film,
punch cards, or punched paper tape.
The machine is likely to be able to put out information on magnetic
tape at the same high speed of 60,000 units of information per
second or 10,000 characters per second. But the best printing speed
of an electric typewriter is about 10 or 12 characters a second.
Card-punching speed is about 130 characters a second. Punch-card
tabulator speed can reach a maximum of about 200 characters a second.
Thus we see that here, too, we may be snowed under with the information
that the giant brain puts out, if we fail to ask the giant only for
what we really want.
MECHANICAL BRAINS UNDER CONSTRUCTION
This chapter would not be complete without mention of the great
mechanical brains that were actually under construction at the end of
1947. In power they are intermediate between the machinery now being
designed, described in this chapter, and the earlier machines described
in the previous chapters of this book.
The mechanical brains under construction on December 31, 1947, were:
Harvard’s Sequence-Controlled Relay Calculator _Mark
II_, constructed at the Harvard Computation
Laboratory, tested there July 1947 to January 1948,
and delivered to the Naval Proving Ground, Dahlgren,
Va., in 1948.
The _IBM Selective-Sequence Electronic
Calculator_, constructed in the IBM laboratories,
Endicott, N. Y., and installed in 1947 at the office
of International Business Machines, 590 Madison Ave.,
New York, N. Y.
Moore School of Electrical Engineering’s _EDVAC_
(Electronic Digital Variable Automatic Computer)
being constructed partly at Moore School and partly
elsewhere, and to be delivered to the Ballistic
Research Laboratories, Aberdeen, Md.
Harvard’s Sequence-Controlled Electronic Calculator
_Mark III_, being constructed at the Harvard
Computation Laboratory, and to be delivered to the
Naval Proving Ground, Dahlgren, Va.
We shall cover briefly (and perhaps a little technically) some of the
main features of the first two of these machines; for, during 1948,
they began to do problems. The other two had not been finished by
the end of 1948 and so would be difficult to describe correctly, for
mechanical brains _grow_, and design changes go on until they are
finished—and even afterwards.
Some information about these machines can be obtained from the
organizations referred to above and from reports that should appear
from time to time in some of the journals mentioned in Supplement
3. There is also a regular section entitled “Automatic Computing
Machinery” in the quarterly _Mathematical Tables and Other Aids to
Computation_, where it is likely that current information may be found.
Harvard’s Mark II
The Harvard Sequence-Controlled Calculator Mark II began to do problems
under test during July 1947. This machine is at least twelve times as
powerful as Mark I (see Chapter 6) and was constructed entirely by
the Harvard Computation Laboratory. The machine contains about 13,000
relays of a new type that will operate reliably within ¹/₁₀₀ of a
second.
Numbers in the machine are regularly of 10 decimal digits between
1.000,000,000 and 9.999,999,999, inclusive, multiplied by a power of 10
between 1,000,000,000,000,000 and 0.000,000,000,000,001, inclusive.
For storage of numbers, the machine has 100 relay registers totaling
about 1200 decimal digits. Also, it can consult any one of 8 tape feeds
for numbers and any one of 4 tape feeds for instructions. Effectively,
the machine can read one number and one instruction from paper tape in
¹/₃₀ of a second.
The machine performs all arithmetical and most logical operations.
In every second it can carry out 4 multiplications, 8 additions (or
subtractions), and 12 transfers. Division is performed by rapid
approximation using the other operations.
In each second the machine can perform 30 instructions. An instruction
is expressed by 6 digits between 0 and 7 which you can select and, in
effect, by 3 more digits fixed by the time (within the second) when the
machine reads the instruction. For example, in the 9th instruction of
the 30 instructions in each second, we can specify a multiplicand. But,
if we do not want to multiply right then—a rare event if we are coding
wisely—we leave the 9th instruction empty. The machine may operate as a
whole, attending to one problem; or the machine may be separated into
halves, and each half will attend to its own problem.
The IBM Selective-Sequence Electronic Calculator
The IBM Selective-Sequence Electronic Calculator was announced publicly
on January 27, 1948, after some months of trial running. It is a large
and powerful mechanical brain, and it is the intention of International
Business Machines to devote it to solving scientific problems. The
staff of the Watson Scientific Computing Laboratory in New York will be
mainly in charge of the machine.
The machine contains about 12,500 electronic tubes and about 21,500
relays. Numbers in the machine are regularly of either 14 or 19
decimal digits. Instructions are expressed as numbers. For storage of
information, the machine has a capacity of 8 registers totaling 160
decimal digits of very rapid memory in electronic tubes. Also, it has
about 150 registers totaling 3000 decimal digits of less rapid memory
in relays. Also, it can consult any one of 66 paper tape feeds; each
row on a paper tape can hold up to 78 punched holes or 19 decimal
digits, and the machine can consult 25 rows on one tape in one second.
These paper tapes together give the machine about 400,000 decimal
digits of memory.
For arithmetical and logical operations, the machine has an
arithmetical unit using electronic tubes. This unit can carry out about
50 multiplications or about 250 additions per second, including the
transfers of numbers. In each second the machine can read and perform
50 instructions, and each instruction consists, usually, of getting
two numbers out of two relay registers, performing an operation, and
putting the result into a third relay register.
Eckert-Mauchly’s Binac
As this book went to press, another mechanical brain, the Electronic
Binary Automatic Computer, or BINAC, was announced on August 22,
1949. This machine was constructed by the Eckert-Mauchly Computer
Corporation, Philadelphia, Pa., for Northrop Aircraft, Inc., Hawthorne,
Calif.
This machine has some remarkable properties. It does addition or
subtraction at the rate of 3500 per second. It does multiplication or
division at the rate of 1000 per second. The input is from a keyboard
or magnetic tape; the output is to magnetic tape or an electric
typewriter. Binac has 512 registers of very rapid memory in mercury
tanks, and each register holds 30 binary digits. The machine actually
is a pair of twins: the storage, the computing element, and the control
are double, and each twin runs in step with the other and checks
every operation of the other. In tests in July the machine ran over
10 consecutive hours with no error. Each twin has only 700 electronic
tubes. Binac handles all numbers in binary notation, except that the
keyboard and the typewriter express numbers in _octal notation_ (see
Supplement 2). Finally, Binac is only 5 feet high, 4 feet long, and one
foot wide.