|
Annals of the History of Computing, Volume 5, Number 3, July 1983 . p253 |
|
|
Annals of the History of
Computing, Volume 5, Number 3, July 1983 . 253 The Making of Colossus ALLEN W. M. COOMBS The article describes the techniques that evolved
for the design and manufacture of the Colossus machines and their
auxiliaries, invented for the purpose of code breaking during World War 11
and installed at Bletchley Park. Categories and Subject Descriptors: K.2 [History of Computing]‑Colossus, hardware,
people General Terms: Design Additional Key Words and Phrases: cryptology,
Bletchley Park,‑‑‑Heath Robinson 1.
Introduction In
the first of these three papers, T. H. Flowers has described how he was by a
series of lucky chances brought into the field of code breaking, and how his
prewar experience and his ideas on the subject of electronic switching were
thereby enabled to burgeon in the making of what were in some sense the first
electronic computers in the world. In the third paper, W. W. Chandler will
expound the problems of installing and commissioning the machines at
Bletchley Park, against the clock, employing construction units made straight
from the drawing board, and with staff who had to learn their expertise as
they went along. This second paper links the other two; I hope to show how
the team set out to design and build large machines of a totally novel form
in minimal time, inspired simply by the firm belief that this could be done,
that it was the right thing to do, and that by guess or by God we were the
people to do it. A
point is first to be made concerning historical sequence and timing. I was
enrolled in the Colossus team in or about September 1943, having until then
been engaged in frustrating the knavish tricks of His Majesty's enemies (as
we say in our national anthem) in other and rather less interesting ways. At
that time the first Robinson machine was waiting to be made to work prior to
installation, and the first (and unofficial) Colossus, the Mark 1, had been
completely designed and was approaching manufacturing completion in the
laboratory at Dollis Hill (see Flowers's paper). It was to be my duty (1 now
gather) to concentrate on the Robinsons exclusively; however, the very
successful demonstration of the Mark I Colossus in December 1943 changed all
that. From then on my principal work was on the Mark 11. 1 was a member of
the design team, and also in charge, under Flowers, of its detailed
production and manufacture. It follows, therefore, that comments in this
article on the original design of either the Mark 1 or the Robinsons are
inferences; I had no direct hand in them, but 1 feel sure nevertheless that
my deductions are reasonably correct. 2. The
Preexisting State of the Art The
Robinson designers were clearly of the opinion, usual in those days (Flowers
himself would appear to have been the sole exception, an experience not
unfamiliar to him), that whereas valves (vacuum tubes) had certain advantages
in speed and other things, still they were fragile and sensitive devices, to
be eschewed as far as possible. A colleague once said to me, "Valves?
Don't like them. Nasty things. They break." The result of that attitude
is typified by the "moduloT' addition circuit used in Robinson (see
Figure 1). The
switching logic in this circuit was performed by changing the phase of a 25‑kilohertz
supply through 180 degrees in the presence of a signal Z from a photocell in
the bedstead, or leaving it unchanged in the absence of such a signal‑phase
modulation, in fact. A succession of switches driven from different
photocells should produce an output in exact phase or exact counterphase with
the original voltage (depending on the modulo‑2 sum of the photocell
signals) and, provided that each switch has a gain of unity, directly
comparable with it in magnitude. In the diagram, block B is a Hartley
oscillator providing the 25‑kh source voltage to as many loads as may
be required (hence the buffering attenuator network) at a level of about 0.5
volts. Block A is one switching stage of a number in tandem, the required
connections being made by telephone cords (shown dotted). The transformers in
any one A block are carefully and accurately tuned to give no phase change
from input to output, except as may be specified by signal Z, which can
produce a complete reversal by operation on the ring modulator shown. The
gain of the valve circuit is adjusted (by means of feedback resistor R) to be
exactly unity from input to output of the adder when it is properly
terminated. The voltage resulting from several stages of switching is then
compared with the original supply in the comparator; in the case shown, it is
arranged that if the modulo‑2 sum of the input signals is zero (an even
number of phase reversals), a null output is obtained, whereas a sum of one
(an odd number of phase reversals) gives reinforcement of the supply, and
hence a rectified signal at the output. The Robinsons worked at 2000
characters per second, and the photocell output was for about twothirds of a
signal pitch; at 25 kh there were thus about nine cycles of signal ideally
available for detection. The principle behind this circuit would
seem to ha been that each stage of modulo‑2 addition could be obtained
by the use of a single valve together with number of other less variable
though still quite accurately determined components such as resistors,
transformers, rectifiers, and capacitors. The switching log was not performed
by the valves at all, but by t. rectifiers of the ring modulator, the
function of t valves being merely that of providing unity gain p stage. In
theory, the circuit was very ingenious; practice, it was a nightmare. Each
stage had to most carefully adjusted to give exactly correct phase and gain
(involving endless experiments with cathode resistors), and even then, when
they had been set up as accurately as possible, by the time a number of
stages had been put in tandem the output was neither in phase nor in
counterphase, but somewhere in between. The effect was very strange; about
six stages in tandem would work perfectly, but the addition two more‑any
two‑would produce a meaningless phase change. So the whole set of
adjustments had be done again, with precisely the same result. This
particular problem was solved (though that may not be quite the right word)
by a piece of inspired serendipity on the part of Flowers. 1 remember going
him in despair after a fortnight of unavailing effort produce circuits that
would work either singly, sixes, or in more than sixes. His instant solution was, "Change the
frequency." I did. It worked. I still don't know why. Nor, he tells me,
does he. Even
when the circuits were working properly, the was still a major problem
arising out of the components used. It has been shown that ideally about nine
cycles of 25‑kh voltage were available. But the circuits contained
inductances and capacitances as well as nonlinear components, and they were
switched with step‑function voltages; they were therefore prolific
transient producers. Nor was it certain that the signals from the photocell
tape readers would be exactly synchronous (in fact, they weren't), yet within
nine cycles it was necessary that the output signal stabilize sufficiently to
give a firm difference between somethiing and nothing for a strobe pulse to
operate on‑and the whole thing had to restore to zero before the next
signal started. We
achieved it‑mostly by inspired juggling with resistors. But it was
quite clear that noisy and edgy circuits like this would not do for the
higher signal frequency (5000 characters per second) to be used Colossus.
Flowers's ideas culminated in the Mark circuits, which had the following facilities. 1.
Isolation between circuits, when required, provided by buffer valves, not by
resistor networks. This clearly implied a readiness to multiply the number
of valves, contrary to the established practice of the time. 2. No inductances in critical circuits, though
electromagnetic relays in slow‑speed circuits were not precluded. With
only resistors and capacitors, transients are simple to calculate and
control. 3. No representation of signals by bursts of oscillatory
voltage. Instead, signals were transformed as soon as possible after the
bedstead stage into directvoltage swings of considerable magnitude, and
extended into what later became known as "non‑return-to-zero"
state. 4. No exact adjustments of gain. The resistors available
to use were in any case not free of drift in value‑nor the valves of
variation in emission‑so signals were at all stages derived from valves
which were fully "bottomed" (that is, with their anodes down as far
as they could be driven by running into grid current) or else fully cut off
(that is, with no anode current at all). The first of these two states is
achieved most satisfactorily with pentode valves (in which anode current is
dependent within limits more on screen voltage than on anode voltage), which
inclined us to prefer pentodes and enabled us to use the last facility. 5. Use of the suppressor grid as a switching means (see
next section). These ideas, together with the substitution of one tape by
thyratron rings (see Flowers's paper), solved the Robinson difficulties, but
they naturally introduced some of their own. What the difficulties were, and
how we dealt with them, will be treated next. 3. The Designing
of Colossus The major advantage of transformers is that the absolute
voltage level of a signal becomes unimportant; our adoption of exclusively
direct‑current signals meant that we were immediately faced with the
fact that a signal swinging about earth at one stage of switching became
(from the nature of valve circuitry) a signal swinging about a higher voltage
(about +70 volts in our case) at the next. If several stages of switching
were involved, then either the absolute level of voltage had continually to
rise or the level had somehow to be dropped (by direct‑current means)
between stages. In later years, some machines did indeed use the variable‑level
technique, to the extent of having voltage rails ranging over 1200 volts or
more; we deemed ourselves short enough of manpower anyway, and we limited
ourselves to voltage rails between ‑150 and +200 volts. This meant that
the voltagedropping requirement was critical, for a change in level by
resistive potentiometer means also a drop in voltage swing of the signal, and
the available swing was not always big enough to be reduced in this way, the
more so since the circuits had to be secure against the initial tolerance and
drift aging characteristic of the commercial resistors available to us under
wartime conditions and in bulk. We devised a method of dealing with this problem, as will
be shown. At this point, however, it is appropriate to indicate another of
our design standards arising largely out of the resistor limitation. Reduction of voltage swing due to a
single potentiometer network is troublesome enough; reduction in a multiple
network (that is, one in which several sources are connected through
resistors to a single output) is clearly far worse, and we avoided such
circuits in all except one case, that of the "scale ‑of‑five"
counter. Instead, we isolated the signals by what came to be called the
"one‑electrode, one‑job" technique. If two signals were
to be combined in some logical way, they were fed to two separate valves with
a common cathode or anode. Better still, they were fed to one valve, but
separately to the control and suppressor grids, for we had realized that a
pentode valve could be cut off on either, thereby giving a logical
"and" for upward‑going signals and a logical "or"
for downward‑going signals. The suppressor technique was so useful, and
we became so attached to it, that after the war we tended to avoid the double
triodes used by most computer makers for their binary stores, on the grounds
that such '1ong‑tailed pairs" left no separate electrode for
switching purposes (apart from other disadvantages). It has to be admitted,
however, that the suppressor swing to achieve cutoff was much greater than
the control‑grid swing, and the circuits had to be arranged to cope
with this. The
result of our deliberations was a set of standard conditions as follows. 1.
Interstage signals to be carried by pairs of wires, one wire at slightly
higher than earth potential, the other at less than ‑35 volts. The
binary signals zero and one were represented by reversal of the voltage
conditions on the two wires. 2.
Signals to be fed out from buffer valves of low impedance, thus allowing
multiple feed‑out, cutting out "back circuits," and avoiding
the worst effects of wiring capacitance (and therefore delay) on interrack
wiring. 3. Normal switching valves
to be small pentodes (code EF36), with triodes for cathode followers and
heavier‑duty valves for buffered outputs. 4. Heater voltage to be
available at two voltave levels, namely, earth and ‑100 volts. It was
permissible to run the heater and the cathode of any one valve at different
voltage levels provided that the heater was not positive relative to the
cathode, and provided that the difference did not exceed about 60 volts. The standard screen
voltage for a conducting valve was about +80 volts relative to the cathode,
derived through a dropping resistor from +100 volts. This gave a control‑grid
base (that is, cutoff voltage) about ‑6 volts and a suppressor base of
about volts, with an anode‑current capability of at leas milliamperes
when the anode was fully bottomed about 15 volts). Control grids were run into
grid c rent in the conducting condition (which in practice meant that they
stood at ‑0.4 volts relative to cathodes). The next section indicates
the effect of these design criteria on some typical circuits. 4. Some Typical Circuits 4. 1 The
Modulo‑2 Adder It is appropriate to start
with the modulo‑2 ad (Figure 2) since it gives an interesting contrast
to Robinson adder (Figure 1). The essential circuit that given in Flowers's
paper, with a pair of output valves added to satisfy condition 1 in section
3. 7 potentiometer chains, the X's and the Y's, are volved. 4.2 The
Shift Register One stage of shifting, or
"remembering one back,' shown in Figure 3. The idea is that a signal A
on input pair of lines is to be read off at a given instant by the strobe
pulse B (derived from the sprocket output) and recorded on the bistable D
(output while at the same time signal A is itself possibly undergoing change
under the influence of the same strobe pulse B‑hence the delay networks
E, which prevent the new signal at A running right through circuit to the
output. Observe that potentiometer ) somewhat more complex in this case,
requiring tighter control on the component selection and tolerance, a factor
now to be considered. In the two cases cited, it
was found that satisfactory voltage swings could be obtained with resistors
vary ±10 percent from nominal. But most of the resist available to us in 1943
(at least in quantity) were guaranteed only to ±20 percent, and were in any
case known to be subject to 5 percent drift with age (meaning within six
months). Since we had to make the machines at maximum speed using large
numbers of resistors, but without sacrificing reliability, this looked like
an impasse. However, we found that the makers, while not able to work to tighter
tolerances, were nevertheless prepared to send us resistors already sorted
into 5 percent groups, though we would have to accept the groups as they came
off the production lines. Thus resistor No. 28 was nominally 220 kilohms; we
would receive supplies of it in eight marked boxes, box 28a containing
resistors of value from 180 to 190 kilohms, and so on up to box 28j, which
contained resistors of value from 250 to 260 kilohms. With the potentiometers
designed to nominal values, the rule was that the resistors in any one
machine were to be taken from boxes with the same suffix letter (or, failing
that, resistors on one rack, or, failing that, on one plate); choosing the
boxes to be as near nominal as were available, we could be reasonably sure
that the resistor ratios were always provided by ±2.5 percent resistors, and
that even with a subsequent drift of 5 percent the voltages derived from the
potentiometers would never exceed tolerance. In fact, they never did. There
was one exception to the rule just cited, which has already been mentioned:
the scale‑of‑five counter. Here we had to depart from the other
rule (against multiple potentiometers), and we found that the tolerance in
the resistors feeding the valve grids had to be much tighter. A description of
this circuit follows. 4.3 The
Scale‑of‑Five Counter The
ring of five valves constituting the scale‑of‑five counter
(Figure 4) was a pentastable (though 1 don't remember that we ever called it
that). The equilibrium conditions were with any one valve cut off and all the
rest bottomed. It is a major problem to calculate the available grid voltage
swings under all possible adverse conditions of resistor variation with such
complex circuits, and really the only absolutely safe tolerance for the
resistors of the multiple potentiometers was ±1 percent‑with no drift.
A few such resistors were available, but not nearly enough (we required 500
per machine), so we had to make do with 4 percent ranges (±2 percent) of the
best quality we could get. It helped that we used lower‑valued
resistors than usual in the screen circuits (10 kilohms for 22 kilohms),
which had the effect of bottoming the anodes to 12 volts, and, as it
happened, the counters were satisfactory; we were probably saved by
statistical averaging of the variations in each resistor network. Observe the use of suppressor‑grid switching in this circuit.
The counter‑drive pulse (of about 10 microseconds duration) was applied
to all the control grids in parallel and had the effect of driving those
grids positive and the valves into grid current. Four of the valves were of
course already in this condition, however; the remaining onethe one marking
the present state of the counter‑responded by bottoming on its anode,
thereby producing a differentiated negative going pulse at the suppressor of
the next valve in the ring, which cut that valve off. The suppressor pulse
was timed to outlast the drive pulse (40 versus 10 microseconds), so that
when the pulse drive had been removed and equilibrium reestablished, the
counter had moved on by one step only. We
may have been luckier than we realized in this circuit. In postwar times we
tried to use valves with short suppressor grid bases, to spare us the large
voltage swings on such electrodes found necessary in Colossus. Valves of this
sort by that time existed; however, whereas the EF36 was insensitive to the
first 8 volts or so of negative swing on the suppressor, the postwar valves
were not. The effect was that sudden swings of the anode were reflected by
capacitative feedback into the high‑impedance suppressor circuit, and
the valves could go into sustained multivibration. But we had no such
problems with Colossus. 4.4
The “Dropper” Circuit The
output from the bedsteads was no more than 0.5 volts, at earth level. This
raised the only really difficult direct‑current amplification problem
in the machine, for the signal had to be raised to standard magnitude, still
at earth level, and bottoming the amplifier valve was not feasible at such a
low input. Our solution was to use that property of the pentode valve which
preserves a fairly constant anode current over a wide range of anode voltage.
Figure 5 shows valve A having a nearly constant screen voltage and grid bias,
and therefore producing a fixed voltage drop in the anode load, largely
independent of the anode voltage; as a result, the anode voltage changes of
amplifier valve B are reproduced with little loss, but at a lower level, at
the grid of signal‑shaping valve C. It was necessary to adjust bias
resistor D of valve A in order to place the grid voltage at the right level
(valve‑emission variation being mostly responsible for this), but the
adjustment was not unduly sensitive, since the swing was by this time large,
and 1 do not think it gave any trouble, once made. 5. The
Manufacturing of Colossus The
work was planned on a sort of "critical‑path‑analysis"
basis (though we didn't call it that then), with the parts likely to take the
longest time started first‑even before we received official authority
to proceed, as Flowers has said. The circuits were designed by Flowers,
Chandler, Sidney W. Broadhurst, and myself‑1 remember Flowers having
made a rough draft of the projected new machine, then tearing his draft into
pieces and sharing out the bits to us with the instruction to get on with it.
Once completed, the designs were handed over to other engineers whose
function it was to lay out the circuits for assembly on standardized plates.
The assembly and wiring were carried out by technical staff, including some of
the wartime female assistants, first at Dollis Hill and later also at the
Birmingham factory, with one of our engineers (0. G. T. Belcher) acting as
resident supervisor. Racks were wired by small squads under more experienced
skilled workmen, and then the whole lot was sent for assembly and
commissioning at Bletchley. About 70 people of many ranks were involved, of
whom some 15 finally went to work at Bletchley as maintenance staff. All
these spent long hours on work of such a nature that its very purpose could
not be revealed to most of them (one bright spark with a knowledge of radio
carefully drew out the circuits he had been wiring up on a plate, and came to
the conclusion that he had been making a lot of directcurrent amplifiers,
which of course he had), and from June 1944 they had to contend with the
extra hazards of flying bombs and V2 rockets. But the work never stopped; the
job was done and the machines delivered, which reflects great credit on those
concerned, especially (when you come to think of it) on those who didn't even
know what they were doing, and to that extent might naturally have lacked
motivation. There
were, of course, other circuits using less novel forms of switching to
control the slower‑speed parts of the machine. Here again luck was with
us. We were not only the Post Office Research Station, knowledgeable about
teleprinters; we were also the Signalling Group, with a wealth of experience
on automatic telephone exchanges and such devices as electromagnetic relays
and rotary switches, and we had at least one genuine wizard in the person of
Broadhurst. Both he and Flowers were the sort of men who could at the drop of
a hat design highly complicated relay circuits to perform any operation
required‑and get it right first time. We
were in the best Shakespearean tradition a few, a happy few, a band of
brothers. I had at that time just had the great happiness of being accepted
by my future bride (and present wife); to grace the occasion, Broadhurst
solemnly named one of his relays VJB (her initials), this being a relay
having a strong interaction with another named AWMC. Eheu, fugaces, Postume, Posturne, labuntur anni. ["Ah, Postumus,
Postumus, how the fleeting years go by. "‑ED.] 6. And
Lastly So
the designs were completed, and such was our confidence in the new technique
that we rapidly reached the stage where, new requirements having been
formulated or new ideas advanced, we would specify forthwith the new circuits
required in the sure and certain knowledge that they would work. Of such a
nature was the special machine we called Protege, another called Aquarius
(which stored information as charges in small capacitors we continually had
to top up‑whence the name), Salamander (which was never completed,
having been forestalled by the peace), and the attachment called Mighty
Wurlitzer, which can be seen as a sort of cinema organ keyboard near the
foreground Wren in Figure 14 of Flowers's paper. The habit of confidence was
hard to break. After the war, as a member of a totally different team on
totally different work, 1 once blandly suggested that such and such would be
the circuit 1 would use, and that would be all right, wouldn't it? 1 mean,
they wouldn't want a mock‑up or anything, would they? 1 was speedily
disabused. War is one thing; peace is quite another. So
nothing remained but to assemble the machines and make them work. But that
was the department of Bill Chandler, whose paper follows. Annals of the History of Computing, Volume 5, Number 3,
July 1983 . 259 Copyright
1983 by the American Federation of Information Processing Societies, Inc.
Permission to copy without fee all or part of this material is granted
provided that the copies are not made or distributed for direct commercial
advantage, the AFIPS copyright notice and the title of the publication and
its date appear, and notice is given that the copying is by permission of the
American Federation of Information Processing Societies, Inc. To copy
otherwise, or to republish, requires specific permission. Author's Address: Blyth Knoll, Yealmbury
Hill, Yealmpton, Devon, PL8 2JX, England. copyright 1983 AFIPS 0164‑1239/83/030253‑259$01.00/00 |
|
|
Allen W. M. Coombs received B.Sc. and Ph.D. degrees
from Glasgow University in 1932 and 1936. He is an associate of the Royal
College of Science and Technology (now the University of Strathclyde). From 1936 to 1973 he was employed by the British Post Office (now British Telecom) on various projects in
its Research Branch. He spent the last 10 years or so of his employment in
research on intelligent machines, in connection with pattern recognition for
postal mechanization. Coombs writes the following about the present..‑‑‑He
is now blissfully retired to the glorious county of Devonshire, where he and his wife can grow their
own food, make their own wine, sail their own boat, and get thoroughly
engulfed in the maelstrom of English village life, and whence they can make
periodic excursions in foreign travel. And despite the recrudescence of many wounds from his student days,
he still believes that the Almighty created heaven and earth with Rugby Union football in
mind." Acknowledgment is made to the Director of Research of British Telecom for permission to make use of the information in these articles. 0 1983 AFIPS 0164-1239/83/030239-252$01.00100 |
|