Cray Research Inc’s managing director in the UK, Chris Windridge, describes the company’s new 16-CPU Y-MP C90 supercomputer (CI No 1,806) as the world’s fastest. Rating a peak performance of only 16 GFLOPS, however, the machine has to be distinguished from other supercomputers. Cray’s point is that the C90 is a real-world supercomputer, not a high-in-the-sky massively parallel machine that might never attain its theoretical peak performance – and that’s not just a dig at Thinking Machines Corp, whose new CM-5 Sparc-based parallel machine claims to have the capability to achieve a peak TeraFLOPS performance for a mere $320m. Cray’s advice is to pay $30m (GBP20m) for a C90, which is available now and has a solid software base, and bank on a guaranteed sustained performance of 10 GFLOPS – far closer to its theoretical peak, Windridge claims, than 99% of massively parallel systems. While the CM-5 might claim to be feasibly capable of reaching TeraFLOPS, Windridge notes, the machine might consistently only ever churn at 50 GFLOPS. The importance of the C90’s binary-compatibility with the existing Y-MP range, Windridge says, mustn’t be underestimated – the machine can automatically run all existing third-party Y-MP code under Unicos, Cray’s implementation of Unix. For example, explains Cray UK sales manager John Fleming, Finite Element Analysis software specialist MacNeal-Schwendler Corp in Los Angeles has built a company around MSC/Nastran code, which doesn’t run on any of the massively parallel machines. It’s the availability of software that dictates the usefulness of the hardware. One of the orders for the new C90 has come from the European Centre for Medium-range Weather Forcasting in Reading, Berkshire, which opted for Cray over all the currently-available massively parallel systems. So there. That’s why Cray on the surface doesn’t seem too concerned about the threat of the various massively parallel initiatives, such as the European Commission’s Esprit project, GP MIMD, which stands to compete with what Cray is trying to put together in its research labs for first release in 1993. While Cray’s Triton multiple-instruction multiple-data massively parallel machine can’t be code-compatible with the Y-MP range, Cray intends to market its future machines at customers with a Y-MP machine. The two supercomputer lines will be complementary, Windridge explains. Users will have a more general-purpose machine running original Cray code, but will use the Cray Triton to run specialised applications which will benefit from the boost in processing power. An expensive investment, having to buy two Crays, but according to Windridge, most users of massively parallel systems already have another general purpose machine. So Cray is confident that its Y-MP C90 (C90 is an acronym for Cray for the ’90s), is a major breakthrough, a true supercomputer that is here now. The machine currently isn’t scalable, which means Cray has a big gap in its product line between the YM-P 8 with eight processors and boasting 2.67 GFLOPS, and the C90 which has 16 processors and delivers 16 GFLOPS peak. Cray says it might scale down the C90 at a later stage, but at the moment the company’s objective is to up the power of its machines. The C90 isn’t simply a double configuration of the Y-MP 8 though. It is based on 10,000-gate ECL arrays from Motorola Inc, whereas the Y-MP 8 uses 2,500-gate arrays from Motorola. This means greater density per chip, thus twice as much logic on-board and a reduction in distance of signal transmission. Also new is the doubling of the vector pipes. Without requiring any tweaking of the software, some of the larger Cray applications can now operate totally in vector mode. And the C90 has four double-word memory ports per CPU, compared to four single-word memory ports on the Y-MP 8. Memory bandwidth has been speeded up to 250Gbps, from 45Gbps, and there are 256 input-output channels, up from 128. And the bandwidth of Cray’s Very High-Speed Channels, VHISPs, has been increased to 1.8Gbps, from 1Gbps. Meanwhile the Cray-Floating P
oint Systems Inc saga continues – all Windridge would say was that the deal is back on again and negotiations are continuing. Floating Point, with its Sparc-based architecture and unique Unified Application Compiler Technology which enables old and new parallel, scalar and vector applications to run side-by-side (CI No 1,652), obviously has some value to Cray, it’s just a matter of Cray working out what that is, and how it will fit in with the company’s existing products. As to the Digital Equipment Corp-Cray distribution negotiations, these too are still under way – the nature of the discussions is such that DEC will distribute the recently-launched entry-level Cray (CI No 1,787) – strange that DEC would want to distribute a product that competes with its own kit. The talks are expected to be concluded soon. – Sue Norris
Content from our partners
