The drag on sales of massively parallel systems is most often attributed to the lack of a machine that is both hardware- and software-compatible with a large applications base. Now Convex Computer Corp is attempting to crack that barrier with its Exemplar SPP scalable parallel processor unveiled yesterday. Developed in conjunction with Japanese partner, steelmaker NKK Corp, it uses up to 128 100MHz Hewlett-Packard Co Precision Architecture 7100 RISCs and runs a Mach OSF/1 microkernel operating system with a layer providing support for applications running under HP-UX Unix. It also provides source compatibility for applications up on Convex’s C series minisupercomputers. The operating system layer enables user-level and application interfaces to run HP/UX binary executables – that essentially delivers HP-UX middleware, including Motif, X Window, the Open Software Foundation Distributed Computing Environment and commands and libraries on Exemplar. HP-UX applications can run without modification with a re-compile, but need to be re-written if they are to take advantage of the system’s parallelism. The SPP/1000CD – compact design – is a server that scales from two to 16 processors and delivers up to 3.2 GFLOPS performance. The SPP1000/XA – extended architecture – is offered with from eight to 128 CPUs delivering up to 25 GFLOPS, about three times the performance of Convex’s traditional vector supercomputers. The 1000/XA comes a series of connected nodes, or building blocks, each with up to eight processors and 2Gb memory. The SPP1000 includes Convex’s two-level Toroidal Interconnect system which the Richardson, Texas firm says eliminates the bottlenecks that often limit scalability. A first Toroidal Interconnect level connects shared memory, processors and input-output components in the building blocks. The second level connects the nodes, retaining, Convex says, a consistent view of memory for all processors. It also provides CTIcache, an interconnect caching system. Each RISC chip comes with 1Mb data and 1Mb instruction cache. There’s up to 32Gb memory distributed between the nodes, which is accessible to all processors. The input-output system transfers data at up to 4Gbps and supports up to 900Gb disk. Each node has a dedicated 250Mbps access to the input-output system. The SPP1000/CD goes from $145,000 to $750,000 – the SPP1000/XA is from $550,000 to $8m. They’re being sold by Convex and Hewlett from the second half of the year. Convex sees discrete manufacturing, government and defence, computational chemistry, oil exploration and production and biochemical research and development as potential markets. Convex and Hewlett-Packard signed a wide-ranging technology-sharing agreement leading to Exemplar back in March 1992, when Hewlett took a 5% stake in the company. Convex hopes Exemplar, of which it has sold one so far, can revive its flagging fortunes – the firm reported a net loss of $61.3m in 1993, with revenues down 16.7% at $193.1m. Convex claims an installed base of some 1,300 machines and says new C-3 series minisupers will be announced later this year.
