Thinking Machines Corp, the Cambridge, Massachusetts-based manufacturer of single instruction multiple data massively parallel supercomputers, is another company in the running to become the supplier of the world’s first TeraFLOPS supercomputer. The company has reported that it is on schedule to produce results of its research, on which it is collaborating with a consortium of 15 US universities, by 1995. This conflicts with Parsytec GmbH of Aachen, Germany, which in March submitted to the European Commission its proposals for a TeraFLOPS multiple instruction multiple data machine by 1993. The race is on. Meanwhile, Thinking Machines has just brought out its CM-200 supercomputer, designed for 100 concurrent users in network and workstation environments. Thinking Machines claims a peak performance of 40 GFLOPS, a sustainable performance of 21 GFLOPS, and a realistic consistent performance of between 5 and 10 GFLOPS. It runs the company’s new Unix-based CMOST operating system, a general-purpose software environment that integrates debugging, performance analysis and data visualisation into a single environment with an interface based on X Window and Motif protocols. The CM-200 costs $10m for a full 64,000 processor configuration – the processors are Thinking Machine’s own 1-bit chips. Of all the supercomputing architectures, it is easiest to programme for single instruction multiple data massive parallelism. And SIMD processing is cheap. Systems based on this design are usually targeted at very specific applications, since they are the least flexible supercomputing machines. This is because, although the memory resource is distributed across the many processors, as with the multiple instruction design, with the single instruction architecture each scalar processor is allocated only one segment of that memory, and all processors are therefore restricted to executing the same instruction at the same time. The best applications for single instruction machines are therefore ones where a lot of similar data needs to be processed very quickly – Dow Jones & Co in the US, for example, uses a Connection Machine to increase the speed of its historical information system. The CM-200 is compatible with Thinking Machines’ existing Connection Machine range, as used by Mobil Oil Corp, Florida State University, Los Alamos and Minnesota Supercomputer Center. Mobil has tested the new machine and claims that an application running at 14 GFLOPS on the CM-2 system, runs at 21 GFLOPS on the new CM-200. Thinking Machines has also announced a version of CMOST that runs in stand-alone mode on a Sun Microsystems Inc workstation. It is a complete Fortran environment, enabling users to develop and test data parallel Fortran programs on their workstations. This package is aimed at the academic sector for students that don’t have access to a Connection Machine. Something to look out for in the near future is the entry-level CM-200 Hummer machine.
