Colossus - The Greatest Secret in the History of Computing

I think one of Churchill’s biggest mistakes was ordering the destruction of Colossus. They probably saw it only as a military asset, but dismantling that and the whole project probably set back UK (and European) computing and science enormously.


Oh, I’d understood that slightly differently. They understood the power of their statistical analysis techniques and the weaknesses of codes, and wanted to keep an advantage in world trade and world power by being able to decode things. So for example the crypto gear helpfully exported to the remnants of Empire was crackable by the UK. And, I thought I’d read that one Colossus may have made its way to GCHQ… indeed, we see this (with a citation!):

Colossi 11 and 12, along with two replica Tunny machines, were retained, being moved to GCHQ’s new headquarters at Eastcote in April 1946, and again with GCHQ to Cheltenham between 1952 and 1954. One of the Colossi, known as Colossus Blue, was dismantled in 1959; the other in 1960.

So, perhaps a setback for science and engineering, but possibly an advantage in power.

1 Like

I would have preferred science and engineering :grinning:

1 Like

And well, yes, they still kept the edge, that’s what I meant by military asset: it’s military even when it’s peacetime, crypto doubly so. But keeping the tech hidden for 50 years, stupid. (There’s very likely some joint work with the US spooks that still hasn’t seen the light of day.)

Some of it wasn’t hidden or thrown away. I remember an anecdote (perhaps in Turing’s Cathedral) where an early computing group was trying and failing to get their preferred memory technology working. A visiting engineer from England suggested using Williams-Kilburn storage tubes, and said they’d had many years of operational experience with them. Only when the engineer had left did the computing team realize that the years of experience must’ve been during or very soon after WW2.


Regarding memory and secret British WWII developments, the rather implicit way, the concept of stored programs was introduced in the EDVAC paper, always made me wonder, if this was hinting at prior art.

The stored program computer idea is implied in Kurt Gödel’s famous 1931 paper (and Gödel was a close friend of John Von Neumann). It is spelled out even more clearly in Alan Turing’s 1936 paper: while Turing Machines have their programming separate from memory, the Universal Turing Machine stores data and the program of the TM it is emulating in the same tape.

What I was referring to is:
The first notion of a relation between memory and programs is provided in First Draft of a Report on the EDVAC, sect 2.4 (b)

(b) The instructions which govern a complicated problem may constitute a considerable material, particularly so, if the code is circumstantial (which it is in most arrangements). This material must be remembered.

but this isn’t more detailed until sect 14.0, where this is eventually to “be remembered”:

Before we can formulate this code, we must go through some general considerations concerning the functions of CC [central control;NL] and its relation to M [memory;NL].

The orders which are received by CC come from M, i.e. from the same place where the numerical material is stored. (C.f. 2.4 and 12.3 in particular (b).) The content of M consists of minor cycles (c.f. 12.2 and 12.7), hence by the above each minor cycle must contain a distinguishing mark which indicates whether it is a standard number or an order.

That’s it! This I call entering with a whimper rather than a bang. :wink:
(Also, mind that there’s still a categorial distinction between orders and data, even, while both are held in the same memory. A distinction, which was soon dropped, but became of renewed interest, eventually.)
Do we get a theory? No. Do we get considerations regarding feasibility? No. Rather, it’s an implication of the minor cycle architecture (fetch-decode-compute cycle) and its relation to a memory architecture (including a distinction between instruction register and memory buffer register), which isn’t pointed out at that instance. It’s a bit like it should be obvious anyway – to those who are in the know (already?).

Even though Churchill demanded the destruction of the Colossi machines, I think that both Cambridge and Manchester University computer pioneers “hit the ground running” immediately after the end of WWII.

Maurice Wilkes war work had given him to exposure of the mercury delay line stores used in RADAR signal processing, and he was bound to associate with electronic engineers at the Telecomunications Research Establishment, who understood vacuum tube ananlogue and digital circuitry.

The Moore School Lectures of August 1946, attended by Wilkes, was the focal point of post-war computer design and engineering. EDVAC, EDSAC and other notable machines were the outcome of this productive gathering.

Similarly in Manchester, Max Newman (ex Bletchley Park) fouded the Royal Society Computing Machine Laboratory at Manchester and recruited Fred Williams and Tom Kilburn (both formerly at TRE), rapidly got a team established after the war and had the Manchester Baby running it’s first program by June 1948, with Cambridge’s EDSAC running by May 1949.

Although the Colossii were mostly dismantled, and Tommy Flowers was committed elsewhere, I think there was sufficient momentum and sufficient technical personnel available to establish 2 major computing laboratories in the UK.

Despite the secrecy surrounding Colossus, I don’t think the early UK computing pioneers were hindered too much.

1 Like

More (much more) on Colossus, the wartime proto-computer. It was in a sense a configurable counter, seeking statistical hints to decode Lorenz (or Fish) settings.

Looking at the technology, counting very quickly was a challenge:

Thyratrons count units of 1, 2, 4, 8; high speed relays count units of 16, 32, 48, 64; nickel-iron relays count 80, 160, 240, 320, 400, 800, 1200, 1600; standard telephone relays count 2,000, 4,000, 6,000, and 8,000.

Here’s a user-friendly video from Google:

Here’s Computerphile on the topic:

where we hear:

a typical run on Colossus to discover Initial Settings on a pair of wheels might take 10 minutes, something like that. And you’ve got to do that for five different pairs. So, y’know you’re taking about an hour to work out settings, if you didn’t know them already. Standing Orders said: ‘You must never take more than two hours’. If you haven’t got it sorted by then, on the settings, give up [and] go to another message. But then, if you knew the settings but didn’t know the wheel patterns that was a huge amount of effort [that] was needed. In fact Frank Carter reckons 10 hours of Colossus time to establish what the patterns of 1s and 0s were on the wheels. Now you’ve realized why they ended up with 10 Colossi at Bletchley Park.

Here’s another, longer, video:

Colossus, built during World War II, kept secret for more than 30 years. Professor Brian Randell tells the story about how he stumbled across a reference to its existence and eventually led to the UK government lifting the veil of secrecy surrounding this pioneering computer in 1975

The engineering of the machine owes a lot to Tommy Flowers - not least the insistence that valves (vacuum tubes) were needed and would, if never turned off, be sufficiently reliable. But the determination of the mechanism of the Lorenz machine and some of the ways to crack the cypher come from Bill Tutte - in “the outstanding mental feat of the last century” according to this short interview:

Tutte was extremely adherent to the Official Secrets Act, but very late in life did write up some of his recollections of his work on this. One of these appears as an appendix to Copeland’s book on Colossus, which can be found on the web if you look hard enough…

So the deciphering of Fish went on until November 1942. Then there came a black day. From then on, the cipher messages came to us without those helpful indicator letters. They were replaced by a simple number. No doubt a German cipher officer would look up that number in a little book and find 12 letters printed against it. But we did not have that little book.

Production stopped, save for the occasional pair of messages in depth. I suppose some German inspector had examined the process of encipherment and had exclaimed somewhat as follows: ‘Hey, you’re giving those bastards information that you don’t need to give them! I don’t suppose it has done them any good, but it’s wrong in principle. Stop sending the 12 letters!’

(Here I have assumed the inspector to have complete confidence in the security of Fish, believing that otherwise he would have demanded much more. On second thoughts it is easy to imagine that he did demand more but was overruled by his superiors. After the German collapse an anecdote came to BP, I know not how reliable its source. A German mathematician had queried the security of a cipher machine early in the war. An army officer had replied, ‘So what? We’re winning the war, aren’t we?’)

In some such manner did that unknown German gentleman, as judged at BP, set going the Computer Revolution.

This wikipedia page has useful background and it links to the Copeland book and also to these American wartime papers. Where we read:

Daily solutions of Fish messages at G. C. & C. S. reflect a background of British mathematical genius, superb engineering ability, and solid common sense. Each or these has been a necessary factor.

G.C. & C.S. Fish sections are organized on the basis of these parametric equations. Mr. Newman must solve the former, obtaining pseudo plain text D from cipher text Z. Major Tester must solve the latter, obtaining plain text P from pseudo cipher text D. Operations (1) and (2) are equally important, since no plain text can be read unless both are solved. However we can say that solution of equation (1) has taken the greatest amount of statistical science, and is of great primary interest to Arlington Hall. Equation (2) has taken much knowledge of language and is of great interest as an art.

Some of the technology:

Thyratrons count units of 1, 2, 4, 8; high speed relays count units of 16, 32, 48, 64; nickel-iron relays count 80, 160, 240, 320, 400, 800, 1200, 1600; standard telephone relays (“P.O. 3,000”) count 2,000, 4,000, 6,000, and 8,000.

Mrs. Miles III has approximately 70 vacuum tubes instead of relays, to do the summing job with less trouble.

Colossus is described on page 108:

COLOSSUS: Bedstead has positions for two tapes, but only one runs at once. Each tape travels past only one set of photocells. Tape is driven by friction between tape and the wheels, without the use of sprocket-pull as in Robinson. Colossus has vacuum tube: chi pattern generator, psi pattern generator, and motor pattern generator; also has a vacuum tube memory so it can delta. “Triggers” (rotary telephone switches in this case) select any of 5 sets of wheel patterns for chi, psi, and motor wheels, each set of which my be plugged up at will and selected at will. There are five counters, which may all work at once if desired, and when counters are overloaded the machine runs idly (without stepping the wheel-patterns) until the counters print out and are clear again. “Span counters” permit the selection of a stretch or span of tape for study, eliminating all before and after the counting.

The Newmanry History also handy. Page 54 of the pdf gives us

A brief description of the two machines has already been given [15(b)]. The essential difference between them is that on Robinson all streams of letters are on tapes. On Colossus only Z is on a tape, the wheels being set up electrically.

As a direct result of experience with Heath Robinson all the improvements needed to remedy these defects (except spanning, whose value was overlooked till later) were incorporated by stages in Old Robinson and Super Robinson, and incorporated at the outset in Colossus.

Operation was not very simple because of the lack of symmetry, accentuated by changes introduced without correcting the “signwriting” on the machine.

Colossus 2 possessed from the first, quintuple testing, a generous switch panel (including not-not), a versatile plug-panel, spanning, a double bedstead, and a greatly increased simplicity of operation.

Intolerable delays and mistakes during wheel-breaking were caused by the need for setting up pins at the back of Colossus and complaints finally extorted the wheel-breaking panel on the front of some machines.

The clamour for specialised gadgets continues : the objection to it is the difficulty of maintaining Colossi unless they are all alike : a device worth fitting to all Colossi is much more welcome.

Colossus soon replaced Robinson for setting and breaking, but Robinson remained indispensable for crib runs in which two tapes (derived from Z and P), must be compared in all positions. A successful crib run usually produced key of such length that wheelbreaking was extremely easy. For this reason four Super-Robinsons were ordered to overcome some of the handi-caps which persisted on old Robinson, and to include spanning whose value had been proved on Colossus.

Hat tip to @Maurici_Carbo for prompting my investigation.


In the 2000s, there was a Colossus-on-a-chip project, concluding – if I recall this correctly – that the theoretical speed was about half of PENTIUM II speed, then the canonical unit for cryptography metrics. (The speed was actually controlled by the sprocket wheel of the paper tape mechanism and this was limited to all practical purpose to a speed, where the input tape would just smolder and maybe produce a bit of smoke, but would not combust due to the friction of the mechanism.)

Said recreation project was heavily based on the US liaison reports to fill any gaps. Sadly, I can’t find a trace of the paper on Google. (Maybe, I’ll find some time to investigate some older hard drives for this.)

One of the innovations of Colossus from the predecessor Heath Robinson, as I understand it, is that the sprocket holes are also read optically (as are two other positions, as well as the official punched holes.) The tapes are driven by large diameter pulley wheels. Colossus has a pair of tape loops, so a new one can be loaded while the other is running - only one loop is active at a time. Although a paper cut at 30mph would be a bit of a hazard.

1 Like

Also, I think, the join of the tape loop was the real breaking point.

1 Like

However, I’m quite shocked that there is no trace of that paper (not even on Google Scholar). It was one of the first papers to come up publicly with technical information on the internals.

We do find that comparison in this short and accessible paper (by Tony Sale):

Colossus is not a stored programme computer. It is hard wired and switch programmed, just like ENIAC. Because of its parallel nature it is very fast, even by today’s standards. The intercepted message punched onto ordinary teleprinter paper tape is read at 5,000 characters per second. The sprocket holes down the middle of the tape are read to form the clock for the whole machine. This avoids any synchronisation problems, whatever the speed of the tape, that’s the speed of Colossus. Tommy Flowers once wound up the paper tape drive motor to see what happened. At 9,600 characters per second the tape burst and flew all over the room at 60 mph! It was decided that 5,000 cps was a safe speed.
At 5,000 cps the interval between sprocket holes is 200 microsecs. In this time Colossus will do up to 100 Boolean calculations simultaneously on each of the five tape channels and across a five character matrix. The gate delay time is 1.2 microsecs which is quite remarkable for very ordinary valves. It demonstrates the design skills of Tommy Flowers and Allen Coombs who re-engineered most of the Mk2 Colossus.
Colossus is so fast and parallel that a modern Pentium PC programmed to do the same code breaking task takes twice as long as Colossus to achieve a result!


Notably, it’s – according to this article – double Pentium speed, at the “safe maximum” of 5,000 cps / 60 mph!

Here’s Bill Tutte’s Fish and I - just 9 pages.

The critic must point out two grave errors, first the poor Ψpatterns and second the sending of a long depth of two. Either error without the other the Germans would I think have got away with. But the two together gave away the structure of the machine.

It was even found possible to break the wheel-patterns for a month from indicators alone exploiting stereotyped beginnings and information from indicators as to which wheels in which messages had the same setting. I remember trying this method myself getting some initial success but soon losing control. Then Capt JM Wyllie tried. In civil life he edited the Oxford Latin Dictionary. "This is just the job for a lexicographer" quoth he. And he broke the wheel patterns for a past month hitherto untouched.

Tutte’s later document “My Work at Bletchley Park” can be read on this page:

I remember an interview with my tutor, Patrick Duff, when he told me that I should go to a certain town about 50 miles away for an interview about a possible war job…

I went into Gerry Morgan’s office to tell of these results. Max Newman was there. They began to tell me, enthusiastically, about the current state of their own investigations. When I had an opportunity to speak I said, rather brashly, ‘Now my method is much simpler.’ They demanded a description. I must say they were rapidly converted. The Research Section urged the adoption of the ‘statistical method’ of wheel setting. Soon the electrical engineers produced the necessary machines. We called them Heath Robinsons. The engineers were pleased because they were rushing teleprinter tape through the machinery at unprecedented speeds without ever breaking it. Well, hardly ever. Agreements, I was told, were counted photoelectrically.

The statistical method worked best on long messages. Fortunately the Germans were sending longer and longer messages at this time. They would put several actual messages into a single transmission. From our point of view that would be a single very, very long message. Production increased again. But somewhere around this time, the Germans decided to change their wheel patterns every day.

I remember being introduced to Colossus. With other members of the Research Section, I was taken to a large room, where a large box-shaped object, sheathed in sheet metal, stood upon a wet floor. If it was 16 × 3 × 5 in feet, that would not contradict my memory. ‘That,’ we were told, ‘is Colossus.’ Gerry Morgan, gazing at the wet floor, remarked that it had not been house-trained yet. We were told that those valves generated heat and the apparatus had to be water-cooled. Alas, there was some leakage. One of my memories of early Colossus-time is still vivid. It is of Max Newman exclaiming with an air of surprise, ‘You know, this thing could do logical operations!’

I see a couple of interesting (and moderately recent) papers by Benjamin Wells on the topic of Colossus, including the important finding “that a universal Turing machine could have been implemented on a cluster of the ten Colossi, proving the power of Colossus.”

There’s also an abstract of one of his talks here:
The Architecture of Colossus, the first PC
but the video of it eludes me. (via)

1 Like