Richardson, Texas-based Convex Computer Corp’s next-generation line of vector minisupercomputers, the C4/XA series (CI No 2,436), is claimed to include the fastest single-processor air-cooled computer built, with the Gallium Arsenide uniprocessor claimed to deliver over 1 GFLOPS at a $750,000 base price with 250Mb of memory. A four-processor combination will cost up to $2.7m with […]
Richardson, Texas-based Convex Computer Corp’s next-generation line of vector minisupercomputers, the C4/XA series (CI No 2,436), is claimed to include the fastest single-processor air-cooled computer built, with the Gallium Arsenide uniprocessor claimed to deliver over 1 GFLOPS at a $750,000 base price with 250Mb of memory. A four-processor combination will cost up to $2.7m with 2Gb RAM. The uniprocessor is about half the price of the C3 model it succeeds and can run up to six times faster, and the company hopes to upgrade most of its 1,300 C2 and C3 user sites to it. It says three are out already with users – it has some 1,300 C series customers and over 1,300 applications up on the line. The air-cooled CPU uses 350,000-gate arrays and performs at a claimed 1.4 GFLOPS running ConvexOS 11.0 Unix. Although C4 and its predecessors use a totally different switching system from the company’s massively parallel Hewlett-Packard Precision Architecture RISC Exemplar system – which offers 4Gbps throughput compared with C4’s 1.2Gbps – Convex plans to merge as much of the technology in the two systems as possible over time, aiming to make them as seamless as possible as far as users are concerned. The two systems already share compilers and C series applications will run on the Exemplar’s SPP-UX, HP-UX-compatible operating system with a recompile. Nevertheless, Convex expects to continue iterating the GaAs line as long as customers demand it. Some are simply not ready for the move over to parallel architectures, the company says. They may be dependent on code from third party vendors that have no plans to parallelise their software, or have written their own custom software specifically for the C series environment. In any case, vector processing still has a place, the firm argues, because of its performance and suitability for certain types of applications, although masssively parallel RISC technology is gradually closing the gap, it says. Exemplar, which is still in beta testing, is now up to 710 MFLOPS in Linpack 1000 by 1000 testing in an eight-way arrangement.