| Machine type | RISC-based distributed-memory multi-processor |
|---|---|
| Models | SCALI |
| Operating system | SunOS (Suns Unix variant) |
| Connection structure | SMP-nodes connected by ring (see remarks) |
| Compilers | Fortran 7790, C and C++ |
System parameters:
| Model | SCALI |
|---|---|
| Clock cycle | 6.7 ns |
| Theor. peak performance | |
| Per Proc. (64-bits) | 150 Mflop/s |
| 64-bits precision | 76.8 Gflop/s |
| Main memory | <=16 GB |
| Memory/node | <=0.5 GB |
| Communication bandwidth | 1.2/1 GB/s |
| No. of processors | 1-512 |
Remarks:
The SCALI system is a non-uniform access system: it consists of SMP nodes that can contain up to 4 processors. The configuration discussed here uses HyperSparc 6.7 ns cycle processors but also HyperSparcs of lower speed or SuperSparc-II processors could be used, showing that the architecture is essentially processor-independent. The SMP nodes are connected by Internode Communication Channel (ICC) links based on SCI, with a bandwidth of 1 GB/s. This is hardly less than the internal bandwidth of 1.2 GB/s of the SMP nodes.
The connection between SMP nodes may be a single ring but richer connection structures, like a ring of rings, a multi-level crossbar are possible for those willing to pay for it.\\ The available compilers provided by Applied Parallel Research are: Fortran 77 and Fortran 90, C, and C++, with facilities for automatic parallelisation. Kongsberg Informasjonskontroll provides a proprietary MPI implementation, MPI-I, for its systems. In addition the KAP preprocessor is available.
In the USA the same system is marketed under the name AxilSCC by Axil.
Measured Performances: The SCALI system is very new. No performance figures are available at this moment.