Jump Over Left Menu
These are a set of papers belonging to Jack Howlett concerned with the potential purchase of a Ferranti Atlas Computer.
10.06.59: The Ferranti ORION and ATLAS Computers, Jack Howlett
Ferranti Ltd have embarked on two new computer projects: ORION, comparable with IBM's 709 and ATLAS, comparable with STRETCH. Both are transistor machines and Ferranti's emphasize that both are to be regarded as computing systems, that is, they are planned to allow for ease of attachment of large numbers of peripheral units such as card or tape readers and punches, printers and magnetic tape decks.
This is entirely a Ferranti design, with the order code largely influenced by their PEGASUS experience. The basic frequency is 500 kcs/sec; fast storage is on cores and can go up to 16384 words of 48 binary digits each. The order-code is 3 address, providing both fixed and floating point arithmetic, and can accommodate 128 instructions of which about 100 have so far been allocated; any of the first 255 registers can be used for B-modification. Times are not yet settled, but the following are not likely to be far out:
|Simple fixed-point arithmetic, logical operations, jumps||60-80|
|Floating point addition (3 address z= x + y)||120|
|Floating point multiplication (3 address z= xy)||200|
|Floating point division (3 address z= x/y)||700|
Ampex magnetic tape units will be used holding about 3 × 106 words per reel and transferring at 11,500 words per second; cad equipment and printer will be by ICT, based on that developed for MERCURY. All transfers will be autonomous, with lock-outs to prevent use by the program of those parts of the store involved in the transfer; this will make time-sharing possible, with a priority list for the different programs. It will be possible to consult the machine at any moment by means of a directly-coupled typewriter.
Ferranti's estimate that, disregarding any effects of the larger fast store, ORION will be 3 to 5 times as fast as MERCURY - this is comparing programs written in machine code, not Autocode. They will develop an Autocode for ORION, and also will program it to accept MERCURY Autocode.
Production of the logical packages and machine frames has started at Manchester, and of the magnetic tape control will start in August in the Edinburgh factory. The first machine, which will be retained by Ferranti's, is scheduled to be working, without its magnetic tape, in June 1960, the tapes to be working in October 1960, June 1961 is given as the date when the first customer would have his machine.
The cost would depend on the amount of equipment; a large installation would cost about £250,000.
Firm figures for specification, cost and delivery should be available about August.
This is essentially the machine designed by Dr Kilburn at Manchester University; it is intended to work at the highest speed practicable with existing components and to provide a computing system of up to very great size.
The machine works asynchronously. The word length is 48 binary digits, made up of 10 for the function code, two groups of 7 for modifier addresses and 24 for the store address. There is a fixed store of 8192 words and access time is 0.2µs which is not accessible to the programmer, but which will hold standard programs and useful constants; the B-store of 128 modifier registers is a core store, with access time 0.5µs. The main store is on cores with access time at most 2µs (there is hope of making this 1µs) and could be up to 224 words; the minimum size will be 213 (8192) words. Present values for operating times are:
|Floating point addition (1 address||1.5-2.0|
|Floating point multiplication||3-4|
|Floating point division||10-15|
Because of overlapping of orders, much of the organisational work in a program will in effect take no time.
Magnetic tape will be as for ORION; all operations will be carried out by routines in the fixed store, which place the initiating instructions in a queue of capacity 64 instructions and lock out the relevant parts of the main store. The main program will continue without interruption unless the queue is full or an attempt is made to refer to a locked-out part of the store. Up to 16 tapes can be reading or writing simultaneously (giving a transfer rate of 180,000 words per second) and at the same time up to 32 others can be searching.
Time-sharing will be possible. It can be arranged that whenever the main program has to wait for some peripheral equipment or a part of the store to become free, a subsidiary program is taken up; thus it will be possible to interleave testing with production, or to use the computer as an off-line printer without loss of production time. As with ORION, it will be possible to interrupt or consult from the console.
Work is going on in Dr Kilburn's laboratory, where at the moment about 5 Ferranti engineers have joined the university team. The plan is to have enough of the machine built by the end of this year to establish the operating speeds and to have the prototype working by June 1962: this is to be a complete and engineered machine with 8192 words of core store, about 100,000 words on drums, and several magnetic tapes. Work on production models will go on in the Ferrant factory in parallel and delivery in 1963 is being suggested. The cost would depend on the amount of peripheral equipment and the size of the core store (this costs about £8 to £10 per word; a large installation would cost between £500,000 and £1,000,000.
15.10.59 W M Lomer to B F J Schenlaml, AERE
A recent meeting of the Computer Policy Committee, chaired in Sir William Penney's absence by Dr Dr Dunworth, came to the conclusion that the Research and Industrial Groups need considerable growth in their computer power if we are to hold our own in reactor design or in CTR experimental design. The availability of Weapons Group machines cannot be guaranteed.
It was suggested that a very large computer - a Ferranti Atlas - should be the mainstay of the increased power. It would be available in early 1963. The needs of the interim period, it was suggested, should be met by the hire or purchase of an IBM 7090. I feel very strongly indeed that we must have these machines if our contribution to the physics of reactors or CTR is to be worthwhile. To get them at the dates we suggest means ordering them very soon.
Howlett, with Iliffe of CTS Risley, has started drafting the case in detail with estimates of cost and staff. These details are nearly impossible to settle without speaking in terms of specific proposals. It is obvious that siting will be a major source of contention. The solution which I would favour at the moment would be Atlas at Harwell and 7090 at Risley.
Two comparatively minor additions which should be discussed (provided they have no influence on the main decisions) are the transfer of the 70% from Risly to Winfrith and the acquisition of an Orion (Ferranti equivalent of the 709) for joint use of the National Institute, Culham and AERE from some date about late 1961.
Unless we make the main moves now, the initiative in reactor design and CTR theory will go entirely to the Americans and Russians.
It is hard to express the urgency of the matter - I would hope that even if we cannot now even provisionally agree on siting that at least a paper with the above notional sitings (in order to estimate costs properly) could go forward and the orders for Atlas and the 7090 placed. Can we have your approval for this paper to be written up and submitted to you in draft in the near future before Sir William is asked to present it to the Authority on behalf of the Computer Policy Committee.
19.10.59 Meeting in P W Mummery's Office, AERE
A meeting was held in Mr P W Mummery's office to discuss what future computational facilities would be required by the Authority for the solution of reactor physics problems. More specifically it was discussed whether an Atlas computer would be required assuming a 1963 delivery date. The meeting was attended by Dr G N Lance and representatives of HTRD and IPRB (Mr G H Kinchin, Dr D C Leslie, Mr C P Gratton, Mr J G Tyror, Mr C T Chudley and Mr J D MacDougall).
It had been previously agreed in discussion between Mr Mummery, Dr Howlett and Dr Lance that the purchase of an Atlas computer was justified if it would either:
- result in a net saving in the cost nuclear power (ie a saving in nuclear power production costs in excess of the cost of the Atlas time employed in obtaining the saving), or
- give a nett reduction in the cost of the reactor physics experimental research programme.
- In addition, there is a case for Atlas if its use would enable solutions to be obtained to otherwise theoretically intractable safety problems which it is highly undesirable to investigate experimentally but for which answers are urgently needed (the case might rest on Atlas computation being cheaper than 7090 computation for such problems).
The meeting discussed the above points and arrived at the following conclusions:
- It was felt that the use of Atlas could result in a definite saving in the cost of nuclear power, but that to make a numerical estimate of the saving to be expected was not very practicable at this stage. This use of Atlas was not discussed very extensively, as it was understood that Mr Iliffe of the D+E group is looking into the savings that might be obtained. However two specific topics were mentioned. Firstly, it was considered that a machine of the size and speed of Atlas was probably adequate to enable a survey to be made of the effect of fuel element shape on reactivity, thus enabling optimum fuel elements to be designed. Secondly, gagging problems were considered. At present, once a reactor is commissioned, the gagging pattern is effectively fixed. In order that a given reactor should be used as efficiently as possible, an optimum lifetime gagging pattern is required; on present generation computers it is possible to compute gagging patterns, but only with an undesirable small number of spatial points, to compute an optimum lifetime gagging pattern would however seem to be a reasonable problem to ask Atlas to solve. [Gagging problems may be within the province of the CEGB rather than that of the Authority.]
- It was also felt that the use of Atlas could reduce the amount of experimental reactor physics work below that which would otherwise be required. Theoretical reactor physics effort would not be reduced but would be re-orientated to ensure the most efficient use of the computer rather than being engaged in correlation of experiments using semi-empirical theories. Atlas would perform reactor physics calculations based primarily on nuclear physics data; it would however be necessary for this nuclear physics data to be supported by a certain amount of integral data from reactor physics experiments. Thus it was envisaged that for the efficient use of Atlas as a reactor physics tool, supporting nuclear and reactor physics experimental programmes would be required, the former to provide sufficiently accurate basic nuclear data on the materials used in reactors, and the latter to provide crucial experiments to confirm the Atlas computations. This would probably mean an increase in the nuclear physics programme and a decrease in the reactor physics programme; the exponential experiment programme would no longer be engaged on performing sufficient experiments to enable a semi-empirical theory to be formulated but would be concerned in performing experiments on stacks specially arranged to test specific parts of the Atlas calculations, and this should involve a smaller amount of experimental work. [Mr Kinchin was not in complete agreement with the above views] The vital point in justifying the purchase of an Atlas is therefore the demonstration that the saving on integral experiments outweighs the cost of Atlas time plus the increased outlay on basic experiments. The basic experiments will probably be the bigger item, but it can be argued that these experiments should be done, even if the Authority does not buy an Atlas. If this is conceded, the case seems quite clear. That as large and fast a machine as Atlas is required for reactor physics calculations is shown, for example, by the estimate that 4 hours of Atlas time would be required to obtain about 0.1% or 0.2% statistical accuracy in K00 in a Monte Carlo cell calculation. To perform a similar calculation on even the 7090 would clearly take an excessive amount of machine time.
- In the realm of safety, it was considered that Atlas might enable calculations of behaviour in fault conditions to be made for situations which could otherwise only be dealt with by experiments which would take a considerable risk of melting down, or even blowing up a reactor. The difficulty in assessing the usefulness of Atlas in safety problems arises from the behaviour of the reactor once physical changes start to take place in the reactor core; thus Atlas could be employed to survey reactor behaviour during reactor incidents up to the time (if any) at which some physical change (such as melting of fuel elements, or incidence of fire) occurred in the reactor. Such a survey would indeed be useful, but description of reactor behaviour after physical change occurs is often required, it was however felt that Atlas would be unable to cope with such situations, owing to the complexity of the phenomena involved. However it might be possible to replace undesirable reactive experiments in which physical changes occur by a combination of Atlas with non-reactive experiments.
Two general comments may be added. Firstly, Atlas has to perform very little more actual computation than either the 704 or 7090 before the cost of a given calculation performed on Atlas becomes less than the cost of the same calculation performed on either of the other machines. Secondly, if a 7090 is rented for a period before Atlas becomes available, it would be highly desirable for the Atlas programming system to be arranged to be compatible with that of the 7090.