If president Jeff Kalb is true to his word, MasPar Computer Corp is going to be big. How big? Kalb does not dream in half-measures – by the end of the decade we’ll be a billion dollar company, he says. And that’s no mean feat considering the Sunnyvale, California-based minisupercomputer manufacturer turned over somewhere between $5m and $8m last year. Kalb reckons sales will triple this year alone. Undoubtedly the company’s newly-established ties with Digital Equipment Corp are somewhere behind these claims – DEC, which will now market MasPar’s MP-1 single instruction multiple data minisupercomputers as part of its own product range on a worldwide basis (CI No 1,655), is expected to account for one third of revenues this year. But, more than that, MasPar is geared up to start delivering parallel supercomputing to the masses, and in so doing the company is confident that it will run away with the market while old hands at supercomputing Cray Research Inc and Thinking Machines Corp are still caught up with the challenge to build the world’s fastest parallel machines.

Commercial

Whereas today, most supercomputers are used in technical and engineering applications, MasPar’s mission is to conquer the commercial market. The commercial applications for massive parallelism envisaged by MasPar are not existing mainframe transaction processing applications, rather they are new data analysis applications. One third-party software toolkit that has been implemented on the MP-1 is Third Eye Software Inc’s Elexir, which enables the development of text-retrieval systems. Using this facility, Kalb reckons one of the main applications for the MP-1 SIMD system will be in similarity searching, a concept tackled also by fuzzy logic experts. Elexir servers running on the MP-1, says Kalb, can search up to 8m documents per second, and the similarity searching facility enables the system to locate similar documents. Lawyers, for example, can quickly access similar legal cases, doctors for comparable medical cases, and insurance companies for similar claims. Another software offering for the MP-1 which is available now is Reading, Berkshire-based Chisq Ltd’s Table-Maker statistical database software, which enables data-intensive statistical analysis, so that banks and insurance companies can download samples of their databases to analyse what if… scenarios. And airlines can use the system for interactively updating flight schedules should a plane go out of action. These applications, which otherwise would be done manually, could be tackled within 10 to 15 minutes on an MP-1, says Kalb. And you can imagine the cost savings to an airline if a day of subsequent flight delays could be avoided.

By Sue Norris

While research figures estimate that the market for massively parallel machines is worth $170m today, and is growing at between 50% and 100% each year, Kalb points out that these forecasts are based only on technical applications – MasPar believes that, by as early as 1995, the massively parallel supercomputing market will be dominated by commercial applications. Kalb’s reasons for thinking this are founded in customer claims – banker J P Morgan & Co, which is testing out a 4,096-processor MP-1 in mortgage and currency trading applications, reckons the MP-1 could be worth $7m in business gains each week. If such claims stand up, then perhaps there really is a case for commercial parallel supercomputing. To ensure that this market takes off, MasPar is making every effort to see that plenty of third-party software becomes available. MasPar’s close alliance with DEC will be instrumental in this process, since MasPar’s resources are limited. In April, when DEC announced that it would take a 5% stake in MasPar and market the MP-1 range, it also unveiled a new business unit to collaborate with MasPar and third-party developers in the development, and conversion of software tools and packages for the MP-1. Kalb says he’ll be happy if there are between eight and 12 strong applications running on the MP-1 in a year’s time. What is important for

MasPar, the president notes with a sly smile, is that unlike the competition, MasPar will not vary its architecture, thus ensuring that its software will always be compatible. In addition to the new MP-1 front-end, the DECstation 5000 Series, software development for the MP-1 can be done on networked Sun Microsystems Inc workstations. MasPar offers two compilers – C and Fortran – and the company reckons that programming for the MP-1 is easier than programming graphics. To refer to the architecture of the MP-1 as SIMD is not strictly correct, since MasPar has its own brand of single-instruction parallelism, which Kalb says combines all the advantages of SIMD design – programmability, debuggability and scalability – with 90% to 95% of the functionality of multiple-instruction frameworks. This has been achieved by introducing four degrees of autonomy to its otherwise sheep-like processors. Firstly, execution autonomy has been achieved by the addition of a switching device on each processor, so that individual processors can decide if they will participate on an instruction. Secondly, processors have the ability to address their own bit of memory.

Simple chips

Thirdly, courtesy of DEC technology, any processor can communicate with any other processor by dialling up a connection – says Kalb, this is more advanced than the message-passing capabilities of MIMD architectures. The fourth degree of freedom is provided by a circuit-switched input-output subsystem, improving throughput. The MP-1 is based on chip technology that has been customised for MasPar by San Jose, California companies Sierra Semiconductor Corp and VLSI Technology Inc. The 4-bit processors, which Kalb acknowledges are not the most up-to-date components, are arranged to give the appearance of a 64-bit architecture – they are assembled in arrays of 32 processors on 1.6 micron chips. The processors each contain either 16Kb or 64Kb of local memory. To get the most out of MasPar’s SIMD architecture, however, requires writing programs specifically to take advantage of data parallelism. Kalb doesn’t perceive this to be a problem, though, since the majority of applications are expected to be new ones. And once the programs have been written, they will run under Unix in a variety of environments. So where is MasPar heading? While the company doesn’t see itself competing in the race for TeraFLOPS performance, the MP-1 is on an upwards climb – Kalb says it should be churning at 200,000 MIPS, 5 GFLOPS by the end of next year.