A string of experimental developments with New York University, Cornell University in upstate New York, and at its own Thomas J Watson Research Center in Yorktown Heights are expected to culminate this year in IBM’s entry into the $2,000m a year scientific supercomputer market – with a parallel machine significantly different from those offered by Cray Research, ETA Systems and the three Japanese contenders. IBM is expected to put its seal of approval on parallel processing with a machine that implements the Hypercube concept of clusters of processors tied together – each node consisting of eight processors, and a maximum configuration having 64 nodes for a total of 512 processors. Conceptually fairly similar to the Inmos Transputer-based T-series machines from Floating Point Systems, in which each Transputer has a Weitek floating point accelerator chip set attached, the IBM machine is expected to come with proprietary 32-bit processors with floating point accelerators attached. IBM has definitely done thorough evaluations of the Transputer, and could be using them in its planned machine. However a key feature of the machine argues against this: it is believed to use an innovative memory architecture that enables the entire main memory of up to 2Gb to be switched around among the processors – and in the limit, the entire memory can be attached to one processor. Such memory access techniques are normally required for handing off tasks from one processor to another, and the architecture of the Transputer is designed to make memory switching unnecessary for that function. Work being done at Cornell on an IBM 3084Q with attached FPS-264 processors from Floating Point Systems may well also feature on the new machine. Cornell has been working on the Gibbs project to develop a free-form near-English langauge that can compile down to Fortran. Such a development could be critical to the future of the new machine, because it will be incompatible with any major machine currently available and there will therefore be no existing base of Fortran applications for it to pick up – although a parallelising Fortran compiler that can take existing Fortran applications and parcel it out for parallel processing would alleviate the problem. Some observers look for the IBM machine to deliver at least 10Gflops and to fall into the $10m to $30m price range. The reason for a high price compared with the Floating Point T-series would likely be that enormous memory, which for speed would have to be constructed entirely out of expensive static RAMs. Within IBM, the supercomputer has gone under the code-name RP-3.
