It's only been just a few short years since the Roadrunner supercomputer broke the petaflop processing performance barrier - that's a quadrillion floating-point operations per second. Shorter still since Fujitsu 's K computer secured the world's top ranking with an incredible performance of 8.162 petaflops. Now, the company has announced global availability for its PRIMEHPC FX10 supercomputer, which can be scaled up to a 1,024 rack configuration for 23.2 petaflops of theoretical processing power.
Petascale computing systems consist of thousands of computing nodes connected via high speed interconnect. Fujitsu's newest system will be made up of 98,304 nodes running on the new SPARC64 IXfx 1.848GHz processor, and will contain 6 petabytes (6,291TB) of high bandwidth memory at its largest potential configuration. Each processor has 16 cores and is capable of standalone performance levels of 236.5 gigaflops, and performance per watt of over 2 gigaflops.
The PRIMEHPC FX10 uses the Tofu interconnect (which stands for Torus Fusion and not the popular bean curd made from soya) - a highly-flexible 6-dimensional Mesh/Torus architecture that has 10 links at each node and is capable of low latency 5GBps bandwidth performance in both directions.
Fujitsu says that the new supercomputer includes the same built-in RAS functions found in mainframes for proven reliability, and benefits from direct water-cooling of components such as processors (although there is also an optional exhaust cooling unit), which reduces overall cooling costs in the equipment room. The company also says that in order to ensure high levels of reliability and operability, all of the components - from processors to original HPC middleware - have been developed in-house.
On the software side, the FX10 operates on a Linux OS with a Technical Computing Suite made up of a high-performance distributed file system called FEFS (capable of sharing across 100,000 nodes) and Fujitsu's own VISIMPACT technology. The latter is said to make automatic implementation of a hybrid parallel programming model a relatively simple affair, enhancing scalability by combining threads - where a process is broken down into multiple threads - with Message Passing Interface (MPI) - which divides a program into multiple processes.
The end user accesses the system via a web-browser-based HPC Portal - to edit files, compile or submit/monitor/kill jobs - while an Operation Management Portal allows admins to monitor and operate the system.
Fujitsu plans to shift at least 50 PRIMEHPC FX10's over the nest three years and sees the system being useful for demanding computer simulation applications like the development of new drugs, disaster prevention and mitigation, advanced product and material development, and the development of new sources of energy. Shipping is set to start in January 2012.