Cambridge, Massachusetts-based Thinking Machines Corp’s new Sparc-based CM-5 supercomputer (CI No 1,792), which is capable of scaling up to a peak performance of 2 TeraFLOPS for tackling grand challenge applications such as wind tunnel modelling, will eventually be able to run software written for an IBM Corp RS/6000, a vectorised ES/9000 or a Sun Microsystems Inc workstation, the company has revealed. This is the goal of a close collaboration with IBM and Sun to develop a common scalable programming environment for scientific computing, probably based on Fortran 90. A timescale for the availability of the applications development environment, though, has not been disclosed. With the CM-5, Thinking Machines is trying to do it all – offer a sustained TeraFLOPS performance to its scientific customers while opening up massive parallelism to the commercial world. IBM mainframe site American Express Co has already ordered two 128-processor CM-5s for use alongside its 3090s, marking Thinking Machines’ second commercial order of all time; the Cambridge company hopes the promise of a common application development standard will start the ball rolling to bring in many more sales based on data processing applications. Meanwhile, Thinking Machines’ development partnership with IBM seems to be more far-reaching than implied. While Thinking Machines is prepared to give little away about the nature of its relationship with IBM, except to says that it will combine Thinking Machines’ expertise in massively parallel supercomputing with IBM’s expertise in high-performance ES/9000 systems, the company at the same time announced that, as from early next year, each processing node on the new CM-5 will carry, in addition to a Sparc chip, four vector units on each processing node, not unlike the ES/9000. The vector option will be available as an upgrade to customers buying the CM-5 before the facility is available. While the old CM-2 Connection Machine, based on custom single-bit chips and Weitek Corp co-processors and scaling down to 2,000 processors for UKP250,000, is based on a SIMD Single Instruction-Multiple Data architecture, which is easier to programme for but less flexible in that it is only suitable for processing large quantities of similar data, the CM-5, which in base configuration has 32 processing nodes – each delivering 128 MFLOPS – and costs UKP1.5m, takes on a new architecture which combines SIMD with the more flexible Multiple Instruction-Multiple Data architecture. The new machine includes synchronisation hardware to induce data parallellism, and CMOST, Thinking Machines’ extended time-sharing version of SunOS, enables different parts of an application to be executed in parallel. This is not real MIMD, since there is no facility for message passing between processors, though Thinking Machines’ UK sales manager Richard Tebboth says the machine can support that level of MIMD processing if required. CM-2 software can be recompiled to run directly on the CM-5, though it’s advisable to tweak the code to gain the benefit from the MIMD parallelism. Once the resources have been ploughed into software development, the CM-5 will support virtual memory. Without this facility, the base configuration CM-5 offers 1Gb memory. A maximum-configuration 16,000-node machine would measure a mammoth 100 feet square, and would have to be put together block by block on the customer’s site. Thinking swears the CM-5 architecture is geared up for this challenge – the communications and throughput channels are claimed to have the capacity to cope with the volumes of data which could be generated by a 16,000 Sparc node machine – it’s just a matter of receiving that monumental order. Thinking Machines has an installed based of 80 supercomputers – 15 in Europe, two in Japan and one in Australia. For the new system, besides American Express, firm orders have been accepted from the Army High Performance Computing Research Centre at the University of Minnesota, which already has its monster machine installed; Schlumberger Ltd; the University of Californ
ia; University of Wisconsin; Syracuse University; Pittsburgh Supercomputing Center; the National Center for Supercomputer Applications; Los Alamos National Laboratory; and Sandia National Laboratories. Most are for systems with hundreds of processor nodes, Los Alamos clocking the largest machine with 1,024 processors. Tebboth notes that US research is much better funded than in the UK, so that US research institutes can afford to install a mixture of Connection Machines and Cray Research Inc supercomputers – therefore, it hasn’t really been a case of making replacement sales.
One-man office
Thinking Machines has its UK base down in Guildford, Surrey, where a one-man office has been in place since February. Similar sales offices are to be found in France, Germany, Italy, Holland, Japan and Australia. At the recent Supercomputing ’91 show staged in Albuquerque, New Mexico, Tebboth reports that although there were some 5,000 exhibitors, the UK was very poorly represented and Australia it seems put in just as strong an appearance.