The SiCortex SC series

Introduction
HPC Architecture
  1. Shared-memory SIMD machines
  2. Distributed-memory SIMD machines
  3. Shared-memory MIMD machines
  4. Distributed-memory MIMD machines
  5. ccNUMA machines
  6. Clusters
  7. Processors
    1. AMD Opteron
    2. IBM POWER5+
    3. IBM BlueGene processors
    4. Intel Itanium 2
    5. Intel Xeon
    6. The MIPS processor
    7. The SPARC processors
  8. Networks
    1. Infiniband
    2. InfiniPath
    3. Myrinet
    4. QsNet
    5. SCI
Available systems
  1. The Bull NovaScale
  2. The C-DAC PARAM Padma
  3. The Cray X1E
  4. The Cray XT3
  5. The Cray XT4
  6. The Cray XMT
  7. The Fujitsu/Siemens M9000
  8. The Fujitsu/Siemens PRIMEQUEST
  9. The Hitachi BladeSymphony
  10. The Hitachi SR11000
  11. The HP Integrity Superdome
  12. The IBM eServer p575
  13. The IBM BlueGene/L&P
  14. The Liquid Computing LiquidIQ
  15. The NEC Express5800/1000
  16. The NEC SX-8
  17. The SGI Altix 4000
  18. The SiCortex SC series
  19. The Sun M9000
Systems disappeared from the list
Systems under development
Glossary
Acknowledgments
References

Machine type RISC-based distributed-memory multi-processor
Models SC648, SC5832
Operating system Linux (Gentoo + extensions)
Connection structure Kautz graph
Compilers Fortran 95, ANSI C, C++
Vendors information Web page http://www.sicortex.com/products
Year of introduction 2006

System parameters:

Model SC648 SC5832
Clock cycle 500 MHz 500 MHz
Theor. peak performance    
Per Proc. (64 bits) 1.0 Gflop/s 1.0 Gflop/s
Maximal 648 Gflop/s 5.8 Tflop/s
No. of processors 648 5832
Main memory ≤ 864 GB ≤ 5.8 TB
Communication bandwidth    
Point-to-point 2 GB/s 2 GB/s
Aggregate peak 648 GB/s 5.8 TB/s

Remarks:

Also SiCortex is a new player on the HPC field, the systems being announced end 2006 and with first available systems in June this year. The systems are unusual in a number of aspects: First, like in the IBM BlueGene systems SiCortex has chosen for a processor with a low clock frequency in order to have low power consumption. According to the documentation this goal has been achieved as the SC648 only uses 2 KW and the SC5832 18 KW. As such they are by far the most energy efficient HPC systems marketed today. The downside is of course the that Theoretical Peak Performance of the processor is also low by today's standards: 1 Gflop/s. The processors are from the MIPS family that disappeared from the HPC scene with the SGI's Origin systems (see Systems disappeared from the list). Now a MIPS64 variant is re-introduced in the SC models (see the MIPS processor for some details). The low clock frequency is not all bad: the mismatch between the memory speed and the processor speed is much smaller than in most other systems which may lead to a higher relative efficiency for codes that suffer a fair proportion of cache misses. Furthermore, the MIPS compilers used to be of very high quality, also potentially leading to high relative efficiency.

The SC systems are built out of 6-processor nodes where all devices of a node are integrated on one chip. This makes the packaging of the systems very compact. A node further contains a common coherent L2 cache for the 6 processors, a DMA engine, two memory controllers, a network fabric switch, and a PCI Express controller. The communication network is logarithmic but unlike other networks employed. It is a so-called Kautz graph which combines a small Ω with a high connectivity, (see [4] and see also the Networks section). The point-to-point bandwidth is stated to be high: 2 GB/s while the latency is extremely low: 1 µs. Because of the constraints of the network we end up with the rather unusual number of processors. In the SC648 108 nodes are connected by a Kautz network with Ω = 4 while in the SC5832 972 nodes are connected with a diameter Ω = 6 network.

Because of the usage of the MIPS processors also MIPS compilers for Fortran and C/C++ are used while a proprietary MPI library is provided. The operating system is based on the Linux Gentoo distribution with some extensions by SiCortex. The file system offered is Lustre as is done by a growing number of vendors (Bull, Cray, Liquid Computing, and many clusters to name some).

Measured Performances:
At this moment no independent performance results are known yet as the first β systems only became available in June this year