The age of TeraFLOPS – nay tens of TeraFLOPS – is upon us with the arrival of Fujitsu Ltd’s new VPP500 series of vector parallel supercomputers, the company claims. According to Stephen Markham, Fujitsu’s Computer Systems Division sales manager for supercomputers, the new machines are truly the next generation of supercomputers. While the nearest competitor, the Cray Research Inc Y-MP C90, offers a configuration of 16 one GigaFLOPS processors, the VPP500 offers a mighty 222 1.6 GigaFLOPS processors (CI No 2,005). What is more, Fujitsu has delivered ahead of expectations. Only last year, the US Department of Energy predicted that 1995 would see the arrival of TeraFLOPS supercomputers. Fujitsu says it will have such machines in the near future, and is now setting its sights on producing 10 TeraFLOPS machines by the end of the decade. The new VPP500 range derives its awesome speed and power from its multiple instruction, multiple data architecture. This enables different jobs to be carried out simultaneously, by different processors, without interruption. The machines’ individual processing elements combine vector and scalar units on multilayer ceramic boards.
Crossbar network
These are connected by a special crossbar network capable of transmitting data at 800Mbytes per second, unimpeded by the kind of jams and interference that dog other parallel configurations, Fujitsu claims. The activities of a VPP500 with up to 19 processing elements are co-ordinated by a 128Mb, 2.1 GFLOPS control processor; configurations upwards of this – to 222 elements, require two control processers. The rationale for combining vector and parallel technology, Fujitsu says, is summed up by (Dr Gene) Amdahl’s Law: When a computer has two different modes of operation, a high speed and a low speed, the overall operation is dominated by the low speed mode, unless the fraction of the work processed in the low speed mode can be virtually eliminated. Some problems contain elements that cannot be parallelised in the conventional way; they require greater and more sophisticated processing power. Fujitsu has opted for a configuration of fewer minisupercomputers, of around a cubic foot in dimension, rather than hordes of cheaper microprocessors in order to tackle such tasks. The company claims that its new system will not only reduce tasks that currently take a year or month to run down to a matter of weeks or days, but ‘will enable scientists and engineers to tackle and solve heretofore unapproachable computational problems’.
By Lynn Stratton
Endorsements have already been received from representatives of key user groups such as the aerospace industry and university academics. For example, Professor Clarke of the University of Manchester Institute of Science & Technology, who is also chairman of Fujitsu’s new Supercomputer User Group, has spoken of the major advances that vector parallel processing would bring to molecular scientific research. The VPP machines are aimed specifically at the high-end of the supercomputing market, at grand challenge scientific and technical markets such as seismic, fluid and structural analysis and simulation; weather forecasting; computational chemistry; drug design and molecular research. They are built entirely from Fujitsu components; even the water for their combined air and water cooling systems is to be bottled and shipped from Japan. They run under Fujitsu’s Unix System V.4-based UXP/VPP operating system, support the Fortran 77 programming language and IEEE data exchange format. Deals with third party developers will be considered, if the demand for application-specific software arises. A roof price for the top end of the range has not been fixed – but UKP5m will buy seven processors and 11.2 GFLOPS, UKP30m will buy 42 processors and 67.2 GFLOPS. Benchmarking will begin early next year with worldwide delivery scheduled for September 1993; the first European recipient will be the Technical University of Darmstadt in Germany. Fujitsu is hoping that it will shift around 50 machines over the next three years;
20 to Japan, 30 overseas (Mr Markham estimates a 50-50 split between the US and Europe). But despite all the excitement, there is no intention of abandoning previous technologies. Tori Minami, managing director of Fujitsu Systems Business (Europe) Ltd says the company will continue to serve the more ‘general purpose’ supercomputing market with its VP2000 and VPX200 vector machines, and will continue to research in the massively parallel scalar direction. Five in the UK Fujitsu, which entered the supercomputing market in 1976, presently claims 23% of the total number of worldwide supercomputer installations with a total of 141 units. It commands 49% of the Japanese market with 106 installations; 60% of these being in the commercial sector, and 40% in the academic sector according to Mr Minami – a reverse of the situation in Europe. At present it claims a 20% share of European installations: 17 in Germany, the result of a fruitful partnership with Siemens Nixdorf Informationssysteme AG; two in France; five in the UK (two at Manchester University, two at Geco-Prakla and one at another oil exploration company); and four others, including a recent installation at the Gallician supercomputing centre in Spain. It has set up a split-site Fujitsu Supercomputing Centre in San Jose, California with an offshoot in Houston, Texas to target the petrochemical and chemical markets, and is especially keen to raise its profile in Europe, which it forecasts will be its biggest market. And this week will see the first meeting of its European Supercomputer Users Group, which intends to meet annually in Europe and Asia.
