nCUBE Corporation
Headquarters 1825 NW 167th Place Beaverton Oregon 97006 USA Tel: (503) 629 5088
Node] --- Based on a 32-bit custom chip and an IEEE-754 compliant 64-bit floating point unit, running at 2 MIPS or 500 kFLOPS (32-bit), 300 kFLOPS (64-bit), with 128 Kbytes of memory. Each node runs the Vertex operating system.
I/O] --- Other I/O boards include: NChannel for interfacing to mass storage devices; a graphics board, with a frame buffer that accepts data from the hypercube and display it at 30 frames/s; an Intersystem board that allowed two cubes to be connected; and an Open System board that allows user-designed circuits to be added.
Topology] --- A hypercube with extra I/O boards, one of which must act as a host board. Host boards can partition the machine into sub-cubes and allocate them separately to different users. They are based on Intel's 80286/7, run the Axis operating system, and can support up to 8 terminals, 4 SMD disc drives and 3 Intel iSBX interfaces. Store-and-forward packet routing is used over high-speed DMA communications channels.
Operating System] --- UNIX or VMS.
Communication Paradigms] --- Systems supported include Express and Strand..
Languages] --- C and Fortran compilers are provided.
Programming Environment] --- Parallel debugging tools, performance monitors and subroutine libraries are provided.
Performance] --- Peak performance is 300 MFLOPS (64-bit) for a 10-dimensional hypercube with 128 Mbytes of memory and an average I/O bandwidth of 90 Mbytes/s.
Scalability] --- Scales from a 4-dimensional cube (16 nodes) to a 10-dimensional cube (1024 nodes).
I/O Other I/O boards include: an improved graphics board; a parallel I/O sub-system board; and an Open System board. Each board can transfer data at a peak rate of 568 Mbytes/s per I/O slot.
Topology A hypercube with extra I/O boards, although host boards are no longer necessary as the nCUBE 2 is hosted by a workstation which is able to allocate sub-cubes separately to different users. Wormhole routing is implemented in hardware over high-speed DMA communications channels.
Operating System UNIX or VMS.
Communication Paradigms Systems supported include Express and Starnd88.
Languages C and Fortran compilers are provided.
Programming Environment The Parallel Software Environment on the nCUBE 2 provides 3 programming models the Node Model, the Node/Process Model and the Host/Node Model --- each appropriate for different types of application. Parallel debugging tools, performance monitors and subroutine libraries are provided.
Performance Peak performance is 27 GFLOPS for a 13-dimensional hypercube with 512 Gbytes of memory and a peak I/O bandwidth of 36 Gbytes/s. It would also be able to support 4096 disc drives. The largest built to date is a 1024 node machine which has been shown to scale almost linearly on some applications.
Scalability Scales from a 5-dimensional cube (32 nodes) to a 13-dimensional cube (8192 nodes).
DARPA appear to have a heavy bias towards Intel at the expense of nCUBE and consequently have always had trouble finding buyers for their machines despite superior quality. However, nCUBE are currently having great success porting Oracle software onto the nCUBE 2, with impressive benchmarks. This may give nCUBE the user-base it needs to compete with Intel and Meiko.