Foster City, California-based nCube Corp, majority-owned by the co-founders of Oracle Corp, surely wins the TeraFLOPS-boasting prize, with the launch next week of the 6.5 TFLOPS nCube 3, its third-generation general-purpose hypercube supercomputer (CI No 1,927). The whopper is scalable to 65,536 proprietary 64-bit processors, boasting an unimaginable peak performance of 6.5 TFLOPS, or 3m MIPS peak performance – 200 times faster than its predecessor, the nCube 2. Also announced is the interim nCube 2S and the entry-level nCube 2E. Due to ship in 1994, with an estimated product turnaround time of 90 days, the monster nCube 3 follows the same architecture as the previous nCube 2 models, enabling software compatibility across the range – the nCube president insists the new machine uses no magic, and claims a full 65,000-processor configuration will be completely feasible – according to Meirer the whole thing would measure just 20 square feet, compared with Thinking Machines Inc’s Sparc-based TeraFLOPS-performance CM-5 which is estimated by Thinking’s Danny Hillis to measure the size of a tennis court in full 16,000-processor configuration (and not 100 square feet, as mistakenly reported in CI No 1,927). A fully-configured nCube 3 would feature 65Tb main memory – 1Gb per processor – with a local memory bandwidth of 500Mbps and an aggregate bandwidth of 32Tbps. The core of the machine will be nCube’s self-contained chip (equipped with DRAM and 40 communications channels), made by Hewlett-Packard Co in Oregon, based on a 3m-transistor processor which has been implemented under 0.6 micron, three-layer metal CMOS technology from VLSI Technology. This will churn some 50 times faster than the version currently used in the nCube 2, with a clock rate that tops 50MHz and a floating point performance of 100 MFLOPS. Machines with 2,000 processors or less will be air-cooled, larger configurations liquid-cooled. The nCube 3 will run a parallelised OSF/1 Unix, and hook up to machines supporting Ethernet TCP/IP, HIPPI, Fiberchannel and FDDI. The interim product, the nCube 2S, meanwhile, will ship this August, offering a 50% performance increase on the existing nCube 2 (which rates a peak 27 GFLOPS), though scalable to the same maximum complement of 8,192 processors, and up to 512Gb main memory. The nCube 2S will be priced as the current 2 series, while the existing range – which runs up to $2m for a maximum 8,000 processors – will be reduced by 35%. No price mark has been suggested for the nCube 3. nCube claims its multi-purpose machines offer a price-performance ratio that is 20 times superior to present-day mainframe computers, making them highly suitable for the commercial data processing market, particularly as Oracle Parallel Server is now shipping with current models. The new entry-level nCube 2E – for parallel software development, entry-level scientific computing or use as a networked database server – is due to ship in October starting at $30,000 in the US, and $40,000 to $50,000 in Europe. It measures 26 high by 18 wide by 30 deep and weighs a mere 150 lb to 300 lb. nCube is currently seeking weighty distribution partners in Japan, the US and Europe; any such marketing agreements will likely be on an OEM basis, the company says. Sue Norris