Cambridge Parallel Processing Inc, the Irvine, California-based massively parallel processing vendor unveiled its fourth generation, Gamma computer at the ACM/IEEE Supercomputing meeting in Washington last week. The Gamma DAP offers up to 1.6 GFLOPS performance, a dramatic improvement on the older DAP 610, and boasts a SuperSparc control processor, increased clock speed and greater than 20,000 peak MIPS in a box the size of a backpack. Cambridge Parallel Processing plans to sell the system in the less than $350,000 bracket. Unlike the majority of massively parallel processing vendors, Cambridge Parallel’s main appeal lies in the high MIPS performance per dollar. When Single Instruction, Multiple Data techniques are used, the biggest problem for the company seems to be in persuading the commercial market of the advantages of truly high speed systems. With around 200 installations around the world, including the recent agreement with the US Department of Energy’s Sandia National Laboratories and as a supplier to Reuters Holdings Plc in London for its newsline service, Reuters Textline, the company is not short on experience. It explains its relatively low profile in the massively parallel processing world as due in part to its involvement in US government classified applications, but that with the demise of the Communist bloc and subsequent US defence budget cuts, a shift to the commercial market has been possible. Clearly unfazed by the recent casualties in the massively parallel processing business, including pioneers Thinking Machines Corp (CI No 2,483), the company believes the variety of applications for which the DAP systems have been used, its legacy of system integration and production readiness, its large base of mature software and the commercial value of its high MIPS performance per dollar will stand it in good stead for the future.
Fine grained
Cambridge Parallel Processing’s use of Single Instruction, Multiple Data bucks a trend away from the technique, although the Gartner Group says market factors that have led some vendors to abandon this technology are based on misunderstanding rather than any shortcomings with the technology. Moreover, few vendors have the time or the resources to develop massively parallel processing and have instead leveraged existing symmetric multiprocessing technology combined into clusters. The DAP Distributed Array Processor, has been around since the early 1970s, in one form or another, first as an Single Instruction, Multiple Data computer system for the British Weather Service and later franchised by the British government to ICL Plc to redesign the system for defence applications. ICL in turn spun off a company, Active Memory Technologies Inc, to market the subsequent commercial version known as the ‘MiniDAP’. Between 1986 and 1992, Active Memory provided a number of upgrades and enhancements to the DAP and in July 1992 the company became US-owned. So to the product: the DAP Gamma is a fine grained, massively parallel processing VMEbus computer containing an array of either 4,096 processors in the Gamma 4000 version, or 1,024 in the Gamma 1000. The elemental processors are custom-built and the machines host Sun Microsystems Inc and Hewlett-Packard Co workstations or network servers. The machines can also be configured as stand-alone computers for dedicated applications. The new Gammas incorporate high-bandwidth ‘Fast input-output’ bus for systems requiring real time graphics display, or for use in ultra high speed data input-output applications such as robotic vision, radar or other sensor based applications. Typical application areas include data mining, complex image processing and data compression.
