Provenance of this document copied by Dai Edwards (dbgedwards@sky.com) on 2/10/2012.
There is no date or author shown on the source document. However, Dai Edwards is sure that this paper was presented at the meeting held on 30th April 1957 at UKAEA. This meeting was called by John Cockcroft. Present were some invited UK computer users and designers (including Kilburn, Wilkes from Cambridge and Strachey from NRDC). The purpose was to discuss technical problems and requirements for a high-speed computer. When the Minutes of the meeting came out, the Three-dimensional computer document was described as 'intended to be read in conjunction with the Minutes'. Dai Edwards suggests that probably Jack Howlett (or possibly E H Cooke-Yarborough) was the author of the paper.
There is a demand for an extremely fast digital computer with a very large storage capacity to perform computation for 3-dimensional field problems. the machine would be requiredto do computations involving a total of about 109 operations on numbers in about 105 addresses. There is room for discussion about what constitutes a reasonable time and what constitutes an operation. The time taken to fill the store from the input organ and to empty it through the output will be considerable, though the need to do this may often be avoidable.
For the present purpose it is convenient to ignore the time taken by the input and output operations and to make the arbitrary assumption that a reasonable time for performing the computation is 10,000 seconds (just under 3 hours). This computing time will be dominated by the time taken to perform a multiplication and this leads to the need for a basic pulse period of the order of 100 millimicroseconds, though values between about 30 and about 300 millimicroseconds can be arrived at by making different assumptions.
Before an operation can be performed, at least one instruction and one or more operands must be brought down from the store. If this process is not to slow up the computing excessively, access to the store for reading should be obtainable at least two or three times per microsecond. This performance cannot be achieved by currently obtainable ferrite materials, so it will be necessary to consider what other materials and what other storage systems might be used. In this connection it is reported that an access time as short as half a microsecond is obtainable with barrier-grid storage tubes. With a diode-capacitor store access time is said to be about 1 microsecond. Thin rolled sheets of magnetic material show turn-over times of about 300 millimicroseconds or less, while evaporated magnetic films are said to be faster still.
Even if the speed of the individual storage element is great enough, it will be necessary to consider whether switching circuits can be made fast enough to select the appropriate one out of all 105 addresses in the available time. An alternative is to have a much smaller fast store backed by a slower, large-capacity store. There will then be the administrative problems of passing information back and forth between the stores. This may complicate programming and, to allow for the extra time taken, it may be necessary to make the arithmetical circuits even faster.
If basic pulse periods in the range 30 to 300 millimicroseconds are considered, it seems likely that pulse delays much greater than 3 and 30 millimicroseconds respectively would be intolerable. This limits the lead lengths to between 60 cms and 6 metres and clearly places a severe limitation on the overall dimensions of the active part of the computer. If the computer can be divided into a number of fairly autonomous units the situation would be improved, but it is doubtful whether this can be taken very far.
The size limitation would appear to make it difficult to use either a large army of barrier-grid tubes or a large diode-capacitor store.
An American paper describes an experimental serial multiplier working with a basic pulse period as short as 10 millimicroseconds. This author, in his conclusions, indicates that he has extreme difficulty in using valves, and looks forward to being able to use transistors. In any case, with valves it would be very difficult to keep down to the necessary small overall dimensions.
For a 100 millimicrosecond basic pulse period, the pulse rise times should preferably be considerably less than 10 millimicroseconds. Thus in a transistor common-emitter pulse amplifier with a current gain of 10, the transistor would need to have an alpha cut-off frequency of at least 40 megacycles, and preferably higher. I.B.M., with a basic pulse period of this order, are proposing using transistors whose alpha cut-off frequency exceeds 200 megacycles. This high cut-off frequency will certainly reduce their circuit problems. We are well behind the Americans in the deveopment of such transistors, but quite good progress is being made, and 100 megacycle transistors have been made experimentally in this country. It should be noted that with graded base transistors the highest cut-off frequency is achieved only at a particular collector voltage and a particular current and it follows that the switching time for such transistors may be considerably longer than might be expected from their small-signal cut-off frequency.
The use of diodes as active elements cannot be ruled out. The U.S. National Bureau of Standards have shown that by making use of carrier storage in the diode the same contact can operate both as emitter and collector. They have made a binary scaling circuit using this principle which will operate at 25 megacycles. While such arrangement is potentially faster than a transistor, since the transit time between emitter and collector is eliminated, the requirement for the driving circuit are very stringent and the circuit problems are therefore formidable. Special diodes can also be used as bistable elements with a very fast switching time. The use of two-terminal active devices does, however, seem to raise difficult circuit problems.
Magnetic cores furnished with a suitable drive waveform can perform as both logical elements and active elements. It seems likely that magnetic material developed for the store might also be suitable for use here, probably in conjunction with transistors.
Diodes are a fairly obvious choice as logical elements, but for high-speed circuits it may be difficult to obtain diodes with a short enough storage time. The forward voltage drop of germanium or silicon diodes may also be large enough to introduce difficulties in circuits where the voltage swings must be reduced to a minimum.
The use of linear transformers to add voltage or currents, with a diode or a transisto used as a non-linear element, may be attractive in high-speed circuits, especially as transformer switching allows best use to be made of the available cut-off frequency of a transistor.
A machine of the high speed proposed would presumably be a parallel one. It appears that the number of basic pulse periods taken to perform a given operation is greatly influenced by the degree of sophistication of the logical design. Thus an ingenious, but perhaps elaborate, logical design might allow the basic pulse period to be lengthened. It seams, however, unwise to rely on this too much, especially in view of the difficulties of circuit technique which are likely to be encountered at high pulse rates. It is probable better to regard logical improvements as a means of regaining speed lost due to unforeseen difficulties.
It seems most unlikely that a problem will ever begin by filling the store with anything of the order of 105 numbers. The store is much more likely to be filled by numbers generated in the course of the computation. Likewise it seems unlikely that the result of a particular computation would be presented as anything of the order of 105 numbers at the output. Some form of direct analogue output might, however, be valuable in some cases.
The whole contents of the store might need to be taken out if they were to be used as a starting point for a further computation later on. Here it would be necessary to balance the time taken to read out the contents of the store and feed them in again later, against the time taken to re-start the computation from the beginning.
The use of punched cards cannot be ruled out, for if the cards are used in most economical way, the time taken to fill the store from the cards would not be out of proportion with the reasonable computing time given above. The use of magnetic tape for the input and output does, however, seem attractive, especially as sorting operations are unlikely to be carried out external to the computer. If the memory of the computer really is large enough to contain all the information needed in computation, then all this information could be fed in from tape in a single operation. this would avoid the need for stopping and starting the tape rapidly under the control of the computer, and should greatly simplify the development of the tape apparatus.
