Fujitsu, leveraging the supercomputing expertise it’s developed through building systems like the powerful K Computer, is now looking to sell such massive computing systems commercially.
Fujitsu officials on Nov. 7 unveiled the PrimeHPC FX10 supercomputer, which they said is capable of scaling up to 23.3 petaflops (or quadrillion floating point operations per second) and comprises technology all developed by the company.
The announcement comes a week before the SC 11 supercomputing show in Seattle, Wash., and about five months after the K Computer-a one-time system built for Japan’s Ministry of Education, Culture, Sports, Science and Technology-was named the world’s fastest supercomputer on the Top500 list. Japanese officials earlier this month announced that that system had hit the 10 petaflop mark.
Now Fujitsu is looking to productize such supercomputing capabilities with the PrimeHPC FX10 supercomputer, which officials said combines high performance and scalability with strong energy efficiency. The company expects to sell about 50 of the systems over the next three years both in Japan and outside that country, noting a growing demand from universities, research institutions and corporations for petaflop-level computing capabilities. The increasingly reliance on computer simulations is helping drive the demand, and the commercial supercomputer will be key in such areas as drug discovery, medical research, analysis of earthquakes and weather, and research into new sources of energy, according to Fujitsu.
Powering the new PrimeHPC FX10 will be the SPARC64 IXfx, Fujitsu’s newest SPARC64 chip that offers 16 cores and a peak performance of 236.5 gigaflops, a much more powerful chip than the eight-core SPARC64 VIIIfx chip that runs the K Computer. The performance-per-watt of the SPARC64 IXfx is more that 2 gigaflops, according to Fujitsu.
The system also leverages Fujitsu’s Tofu technology, an interconnect that offers high memory, bandwidth and scalability with 10 links, with each link offering high performance levels of 5 Gbps in both directions. The PrimeHPC FX10 also will come with Fujitsu’ original high-performance computing (HPC) middleware, called Technical Computing Suite, which includes a compiler and library aimed at ensuring high levels of performance for massively parallel applications.
Scalability also can be enhanced through Fujitsu’s VisImpact technology, which enables customers to create a hybrid parallel programming model that combines threads with Message Passing Interface.
The system can scale up to 1,024 racks, which will contain 98,304 CPUs and 6 petabytes of memory, all delivering a peak performance of up to 23.3 petaflops. Fujitsu’s PrimeHPC FX10 also can come in configurations as small as four racks that offer up to 90.8 teraflops of performance.