Parsytec GmbH reckons that while the $1,650m market for supercomputers is growing at only 2% each year, the market for parallel supercomputing, valued at $150m, is growing at an annual rate of 50%. With 700 installations worldwide, Parsytec claims 20% of the market for multiple instruction multiple data massively parallel supercomputers – matching that of Bristol-based Meiko Scientific Ltd. And that’s pretty impressive, given the size of the company. Six-year-old Parsytec, based in Aachen, Germany, expects sales of $13m for the year to August 1991, up from $9m sales last year. The company, which has subsidiaries in the West Chicago in the US, Pangbourne, Berkshire in the UK, and Chemnitz in the former German Democratic Republic, employs only 130 staff, the majority of which are design or application engineers.
Climate
The company believes, however, that it is the small companies that bear the most fruit – after 30 years waiting for a revolution in power computing from the likes of Bull SA, Siemens AG and ICL Plc, says managing director Falk Kbler, Parsytec was started up to try and effect that longed-for breakthrough. Small compani-es, he explained, have the advantage of being able to concentrate on what they’re good at. In Parsytec’s case this is parallel supercomputing. The company’s new GC range of scalable multiple instruction multiple data super massively parallel computers, announced last Friday (CI No 1,699), have been designed specifically to deliver several hundred GLOPS perfomance, targeted at Grand Challenges in laboratory research and engineering, where vector supercomputing power is not sufficient. The simulation of both global and local atmospheric flow for climate research, for example, is still impossible on even the most powerful supercomputers available today. Cray Research Inc has itself admitted that massively parallel supercomputing is the only means of approaching the TeraFLOPS performance required for running such simulations simultaneously. The Parsytec GC Series is a step in that direction. The company’s vision of the future is such that the super massively parallel machine will co-exist with the vector supercomputer, rather than replace it. The technical realisation of this connection would be via a HIPPI high-performance parallel interface – an ANSI standard interface enabling bulk data to be transfered between a computer’s central processors and input-output devices at very fast rates, accelerating numerically-intensive applications.
By Susan Norris
The new supermassive high-performance GC Series is based on Inmos International Plc’s latest Transputer, the T9000, which uses integrated communication systems to support the interconnection of a large number of nodes. The series has been developed from top to bottom, rather than the other way round, so it is possible to control computers with 16,384 down to 64 processors, the smallest unit. The performance capacity of the GC Series, therefore, ranges from 400 GFLOPS down to 1 GLOPS. The top-of-the-range 400 GFLOPS machines measure a huge 20 square feet, containing 16,384 Transputers, in GigaCubes of 64 processors, each yielding 1 GLOPS. The system’s cost-performance ratio is fixed at $300,000 per GFLOPS, so that in base GigaCube form, the GC with 64 T9000 processors, delivering 1.6 GFLOPS, will cost $500,000. The cluster with 16 T9000, error-correcting memory and four Inmos C104 routing chips, forms the atomic unit on which the GC is built. The 16 T9000s can communicate at a maximum of 1.3Gbps, of which 1.1Gbps can flow through the links to other clusters. And local, autonomous input-output units with a bandwidth of 160Mb can also be connected to each cluster. A seventeenth processor ensures that the system continues operating in the case of a processor fault. For cooling, Parsytec has developed a new system of heatpipes, which keeps the cooling system and the processor modules apart. With this method, heat dissipation can be achieved by air or by water, depending on the size of the computer. As a small company, Parsytec relies on strategi
c technological partnerships – with companies such as Bristol-based Inmos for its Transputers and Amsterdam-based Associated Computer Experts BV, ACE, for its software development environments and compiler technology. ACE’s contribution has been in the development of the Parix Unix-based applications development environment for the GC Series, and in the development of an optimised Fortran 77 compiler – compilers for C, Pascal and Modula-2 have also been adapted for Parix.
Parix
Within the Parix environment, the programmer can develop and run applications on three levels – the Parix light-weight core supplies the basic functions for parallel processing; the advanced programming level adds virtual topologies and global operations; and the third layer provides software support, including numerical libraries, emulation of other parallel processors such as the Intel Corp iPSC, the exploitation of all host functions, and the continued use of existing Transputer software, such as Perihelion Ltd’s Helios CDL and the Inmos Toolset. Parsytec’s new machines are scheduled to be in prototype by the third quarter, and the first systems will ship in a year’s time – six orders have so far been received, for 128 and 256-processor machines, with a total value of $5.5m. Users will include the US Science Support Laboratory in California and the Naval Research Laboratory in Washington. The sales figures from these orders will show in Parsytec’s 1991-92 figures. The company aims to build up to 30 GC systems in 1992, projects $100m sales by 1995. Some $5.6m has been invested in the GC Series, whose architecture forms the basis of Parsytec’s proposal under the European TeraFLOPS Initiative to build a TeraFLOPS machine by 1993 – Parsytec put its proposals forward in March, having been approached by the European Commission at the beginning of the year to take part in the project.
