This week Tokyo Tech announced the winner’s of its competitive process to build the 2.4 PFLOPS successor to its flagship TSUBAME HPC system
The procurement process concluded on May 25, when the NEC-HP partnership’s winning bid was announced. The theoretical maximum performance of the system is 2.4 petaflops, currently the world’s fastest, improving by 30 times the performance of TSUBAME 1.0. The new supercomputer will be 12 times faster than Japan’s current fastest, which is operated by Japan’s National Atomic Energy Agency.
Some of the features of the new system include GPUs and SSDs, which Tokyo Tech hopes will make the machine both efficient and green (the press release indicates they are shooting for a PUE of 1.277)
The TSUBAME 2.0 supercomputer is equipped with cutting-edge technologies such as the latest Intel Westmere-EP and Nehalem-EX processors with “scalar operation,” and will employ approximately 4,200 NVIDIA Fermi GPUs. This “mixed scalar-vector architecture” will achieve world-class computing.
The system has more than 1,400 compute nodes and uses Voltaire’s QDR InfiniBand network. It uses the latest SSD technology and high-density mixed technology for the world’s fastest total data I/O performance at 0.66 terabytes using DataDirect Networks storage technology.
I assume that’s .66 TB per second.
Interestingly, the system will run a mix of Windows HPC Server and Linux, and will use virtualization “to take advantage of the flexibility of cloud hosting services” — thank goodness they managed to wedge “cloud” in there.