The Tokyo Institute of Technology revealed on Wednesday that its next-generation supercomputer will start operation in the fall. Called Tsubame 2.0, it will use 2,816 six-core, 2.96GHz Xeon 5600 processors combined with 4,224 of NVIDIA's Tesla M2050 general-purpose GPUS to top 2.39 petaflops. It should rank second in the Top 500 ranking as of June 2010 and will also be the first petaflop supercomputer in Japan.

Because of its proposed vector-scalar mixture architecture, the computation capacity will differ depending on the type of calculation it must perform. In the Linpack benchmark, the performance is rated at between 1 and 1.4 petaflops (double-precision value), but when faced with vector computer calculations, the performance can be as high as 150 teraflops.

To reach this level of performance, the University improved the network bisection bandwidth compared to Tsubame 1.0 by about 33 times. It is now set at about 200Tbps. Total memory bandwidth is 720Tbps, or 42 times higher than the 2006 original. The computation capacity is improved by a factor of 30.

Multilevel storage made up of DDR3 DRAM and solid state drives was also used. Total memory capacity of the backbone system's DRAMs is 80.6 Tbytes for microprocessors and 12.7 Tbytes for GPUs, while the total memory capacity of the SSDs is 173.9 Tbytes. SSDs have greater performance than traditional hard drives. Tsubame 2.0 also reduced its power consumption versus performance ratio, to the tune of about 1/25.

This efficiency also helps to reduce the cost, which at about $35 million for the build and four years of basic maintenance is said to be low. Construction on the device is yet to start, but will be done by NEC and HP.

A Tsubame 3.0 supercomputer is already said to be starting development, with targets of 30 petaflops and the same or lower power consumption levels of Tsubame 2.0. It is expected to launch in 2014 or 2015.