Illuminata Research is impressed with the new architecture of Hewlett-Packard Co’s V2500 high-end servers, which combines 32- way symmetrical multiprocessing with a cache coherent non-uniform memory scheme able to link up to four of the 32-way nodes together, and beyond that the ability to cluster up to 16 of the 128 processor units into a system with up to 2,048 CPUs (assuming that money is no object). To our eyes, it’s a straightforward adoption of NUMA into its engineers’ toolbox, albeit with much larger building blocks (32 CPUs vs 2 to 4 normally), it says. Illuminata points out that HP has been slower than the competition to scale up its systems, remaining at 16 CPUs while Sun and others moved to 64 processors and beyond. The new announcements sees it catch up, and even go beyond its rivals. Its Convex-derived HyperPlane crossbar switch, and the new 440MHz PA 8500 chip, which equals Compaq’s 600MHz Alpha 21264 in performance at the top end of the CPU charts, give it further advantages, says Illunminata. Silicon Graphics Inc is not so sure. While it says it welcomes HP’s endorsement of the ccNuma concept, it believes latency penalties between each of the 32-way nodes will be too high to run sensible application across boundaries. SGI has implemented a hypercube architecture so that each processor can communicate with other across multiple two-way paths. SGI wonders why HP issued a 64-processor benchmark for Linpack, but only a 32-way test for SPECint – and no SPECfp results at all.
