Sign up for our newsletter
Technology / AI and automation


When Don Gibson, senior engineer for engineering in scientific requirements, breezed into town from the Data Systems division at IBM Poughkeepsie he could be forgiven for feeling a little neglected. The computer press corps had stampeded to the Personal Computer pastures of Greenock and left the man personally responsible for IBM’s 3090 Vector Facility, to proceed with a one-to-one seminar in an auditorium eerily empty as the mid- voyage Marie Celeste. The subject was the 3090 Vector Facility or ‘Optional Frame Five’ – a reference to the extra box on the 3090. In its progress report IBM went as far as to say that a guesstimate of just three sold in the UK was ‘not far off the mark’.

Successful in Japan

Of the three main geographic markets IBM has hit with the Vector Facility, it has been most successful in Japan where 10% of the 3090s installed last year in Japan had Vector Facilities. The Japanese were, said Gibson, further up the learning curve in their use of the Facility than either the US and Europe. This, he said, had much to do with the enlightened funding of industry research by the Ministry of International Trade & Industry and the well-established use of vector applications by Japanese vendors – Hitachi has been offering an embedded vector processor with its IBM-compatible mainframes for a decade now, starting with the M200H which was a contemporary of the IBM 3033. European acceptability was inhibited, he said, by the export licensing problem which was ‘a bit of a headache’. For me the product is 10 years old, said Gibson, In 1977 I began having discussions with large customers in the automobile industry who wanted to use computational models to test the safety of cars. They needed to simulate a physical event and make the computer test objects to destruction. The essence of this work is speed and related to speed is the notion of vectorisation. Before this we had scaler processing where numbers are ‘crunched’ one by one. A vector is simply a intense computing of a string of numbers but IBM has never been known for this, only for very fast scalar processing. There were ‘surprise applications’ outside the more common ones such as computational fluid dynamics and solid state modelling. One Japanese securities firm uses it as a boardroom tool for forecasting the price of stocks if certain investments are made. Each manager can input his choice and get a stock forecast displayed on screen and each of the managers around the boardroom table can share the results in real time in order to get a consensus of which stock to invest in before the market opens. Some new uses include attaching the Nobel Prize-winning tunnelling electron microscope to a Personal Computer and using the Vector Facility to reproduce and display on the micro a picture of an atom in motion at a rate of 30 frames a second. It is also being used for world econometric forecasting by videoconference for the United Nations. Gibson stressed the fact that the Vector Facility began way back with the original 360 architecture and that programs written for the 360 will run unchanged through the 370 and 370XA. The Vector Facility simply adds a vector execution element. A vectorised program has both scalar and vector instructions, the scalar are routed to scalar registers and vector to vector registers. There is full compatiblity so the same data can be processed by either. He highlighted the fact that ‘unlike other vector processors’ the IBM vector architecture allows the process to stop at any number in the ‘string’. The benefit of such precise interrupts is that scalar and vector processing can be intermixed freely. If this was not the case, a vector application, which tends to be very demanding, could lock out the other applications. Addition and logic operations and multiplications and division can be carried out at the same time via two pipelines at a rate of two results per 17.2nS. One vector element can achieve 116 Mflops – floating point operations per second. The maximum six-CPU on the 3090/600 configuration can carry out 700Mflops. IBM

had rated the vector performance at between one and a half and three times scalar performance, but, a year on, Gibson reports that Statoil has found its machine exceeds those expections, while Merck Pharmaceutical achieved the claimed performance, and admitted its expectations had been ‘too high’. We have done hundreds of benchmarks, and most applications fall into the performance range. Sometimes it falls below, in cases where the vector compilation is too short, and then the scalar processor is faster. He gave an example of the Vector Facility in use on a problem showing a three-dimensional air flow pattern between rotary turbine blades. A 3083 took 24 hours, an unenhanced 3090/180E took eight hours, adding a Vector Facility reduced that to three hours, while the top-of-the-range 3090/600 with six Vectors took one hour to produce it. It literally makes weekend jobs into overnight batch, overnight batch into batch and batch jobs virtually interactive claimed Gibson. He added that the Vector Facility had given some of the current users their first taste of parallel processing, and his emphasis of the advantage of combining scalar and vector capabilities in one machine was backed up by Morley Sage of Southampton University. Sage had compared the traditional arrangement of a mainframe attached as a front end to an array processor with the 3090 Vector Facility in terms of speed, ease of use, and range of available software. It was the balance of scalar and vector capability, and the ease of use of the IBM 3090/150 that won against the rather specialised scientific orientation of the traditional approach. That meant that a wider range of students could use the machine for a variety of tasks at a reasonable cost. The users moved from ICL and Honeywell kit to this in about a month and they had to learn how to make use of the vector. There has been no specific training and we have a significant number of users of the vector, said Sage.

White papers from our partners

Doubling throughput

Ease of vectorisation is one of the main factors in the uptake of the system. It may not be as good as a supercomputer for certain tasks but the power is available and usable in a machine doing compute-intensive work with data processing activity. It is not just a powerful numerical machine, and this is important in terms of productivity. The entry level cost of attaching a Vector Facility 13% to 15% of the price of the 3090/150E, which works out at about UKP300,000 – it appears you pay more for the cabinet to start with! On the big systems it approximates 10% of the processor price. As a footnote to price, Bell Helicopter, which invested $5m in a 3090/200 and spent $500,000 to add two Vector Facilities, says the enhancement turned out to be a bargain, doubling the throughput for its Nastran jobs.

This article is from the CBROnline archive: some formatting and images may not be present.

CBR Staff Writer

CBR Online legacy content.