The SC series computers are just Linux clusters - you bring your Linux/MPI program, type "make" and run. Underneath the covers these systems are quite novel - systems designed from the silicon up as clusters, with extreme attention paid to reliability and low power at large scale.
The SC5832 is composed of 972 compute nodes, each of which is a six-way 64-bit SMP running Linux. All node logic, including a high performance cluster interconnect, is contained on a single chip, except for commodity DRAMs. The node chip, implemented in 90nm CMOS, is 1.6 square centimeters, contains about 150 million transistors, and consumes less than 10 watts. The compute nodes are connected via a degree-3 diameter-6 Kautz network constructed from 2916 individual 2 Gigabyte/second point to point links. The interconnect fabric provides low latency, high bisection bandwidth and substantial fault tolerance. Nearly all system software is open source, and based on GNU/Linux.
The talk will cover aspects of system design, with particular attention to architectural balance, low power, reliability, and communications. Some performance results will be presented. As a teaser, the system achieves MPI latencies of less than 1.5 microseconds and an optimized GUPS of about 2.25, placing it 6th worldwide.
Individuals wishing to get a jump start on the talk might look at:
http://www.sicortex.com/products/white_papers/sicortex_technical_summary
http://www.sicortex.com/products/white_papers/a_new_generation_of_cluster_interconnect
