next up previous contents
Next: The NEC Cenju-3. Up: Distributed-memory MIMD systems Previous: The Meiko Computing Surface

The nCUBE 2S.

Machine type Distributed-memory multi-processor
Models nCUBE 2S
Operating system Internal OS transparent to the user, SunOS (Sun's Unix variant) on the front-end system
Connection structure Hypercube
Compilers Extended Fortran 77, ANSI C, C++

System parameters:

Model nCUBE 2S
Clock cycle 40 ns
Theor. peak performance
Per Proc. (64-bits) 3.0 Mflop/s
Maximal(64-bits) 19.7 Gflop/s
Main memory <=256 GB
Memory/node <=64 MB
Communication bandwidth
Point-to-point 2.75 MB/s
No. of processors 8-8192

Remarks:

The nCUBE 2S is, apart from the SGI Origin 2000, presently the only commercially available machine with a hypercube structure. The nCUBE uses in-house developed processors which have a performance of 4.1 Mflop/s in 32-bit precision and of 3.0 Mflop/s in 64-bit precision. The 2S model is virtually identical to the former nCUBE 2 series, however, the clock cycle of the processor has been speeded up by 20%. The communication speed is 2.75 MB/s and a wormhole routing technique is used that makes the communication speed almost independent of the distance between nodes in the system. The maximum size of the nCUBE 2S is 8192 processors.

There are 14 outward connections per node of which 13 are used for inter-processor communication while the fourteenth is used for the distributed I/O system which therefore has the nice property that it scales with the number of nodes.

Within the hypercube sub-cubes can be allocated to accomodate more users. A queue of tasks is set up with (sub)-cubes of the required size. Programs may be written to determine the sub-cube dimensions just before execution.

Some years ago nCUBE had plans to develop a successor to the nCUBE 2S, the nCUBE 3. However, the system was never built after repeated delays and the attempt to build it has been abandoned.

Measured Performances: For a 1024-processor nCUBE 2 a speed of 1.9 Gflop/s was observed for a 21,376-order linear system in 64-bit precision ([4]), about 77% of the quoted peak performance.



next up previous contents
Next: The NEC Cenju-3. Up: Distributed-memory MIMD systems Previous: The Meiko Computing Surface.



Aad van der Steen
Wed Mar 5 09:04:46 MET 1997