Cray users will be pleased to hear that Silicon Graphics Inc has re-affirmed its commitment to the mature but important supercomputer market by unveiling a fourth generation of its CMOS vector processing technology said to be capable of delivering 4 GFLOPS performance per CPU. However to gain the maximum benefit from the additional performance and […]
Cray users will be pleased to hear that Silicon Graphics Inc has re-affirmed its commitment to the mature but important supercomputer market by unveiling a fourth generation of its CMOS vector processing technology said to be capable of delivering 4 GFLOPS performance per CPU. However to gain the maximum benefit from the additional performance and expandability it offers, most users will have to recompile existing applications for it. The new Cray SV1 series will serve as the upgrade route for three of Cray’s four vector lines; the J90, and older YMP and C90 generations, and run the Unicos operating system. A second generation SV2 system not due until after 2000 but suposedly capable of ‘tens of TFLOPS’ performance will succeed both SV1 and the existing high-end T90 system. The T3E continues as a massively parallel Alpha RISC system. SGI’s new CMOS processor includes a vector cache that provides faster access to memory where previous designs included small scalar caches. The 4 GFLOPS-capable processors can be run as four 1 GFLOPS processing elements and each SV1 node can be configured with mixed arrangements of processors that deliver an SMP-type system arrangement with a maximum of 32 GFLOPS. By year-end SGI will offer its promised 32-way superclustering mechanism for configuring SV1 systems with a performance of 1 TFLOPS, double that other vector architectures, it claims. Cache coherent distributed shared memory architecture configurations won’t be possible until SGI supports its core ccNUMA technology in SV2. Until then users running single applications will use message- passing or MQE for multiple applications, SGI says. Its 400 or so J90 users will be able to use existing storage and peripherals with SV1 and although Y-MP and C90 users will have to buy new systems they are promised some degree of application compatibility. The SV1 design will be tweaked to support 1.6 TFLOPS sometime after 2000 while SV2, will support tens of TFLOPS by incorporating MPP functions from T3E, the company claims. SGI’s aim is to build multiple product lines around the ccNUMA technology it curently offers only in its MIPS RISC-based Origin Unix servers. That includes new generations of vector systems and the Origin servers and workstations and new desktops that will in future use 64-bit Intel Corp 64-bit chips. The lines will share a common software architecture. Many data center features of SGI’s Irix Unix have already been incorporated into Unicos and by year-end Irix will feature a so-called ‘supercomputing API’ enabling developers to write applications on Irix systems before moving them to a vector system. Over time the API will be supported on all of SGI’s products, Windows NT included it’s presumed. SV1 deliveries start in August, prices start at $500,000.