While wandering the SC09 show floor on Tuesday, I happened to catch the end of a presentation by Satoshi Matsuoka, the project lead for the TSUBAME supercomputer at the Tokyo Institute of Technology. The current TSUBAME cluster there delivers about 163 peak teraflops, and is sitting at number 57 on the TOP500. It’s made up of an eclectic mix of 655 Sun Fire Opteron blades, 648 ClearSpeed boards, and 170 NVIDIA Tesla servers.
According to Matsuoka, the next generation machine, TSUBAME 2.0, will be a 3 petaflop machine made of next-generation x86 CPUs — not sure if it’s Xeons or Opterons — and NVIDIA Fermi Tesla devices. He thinks the power draw will be in the neighborhood of 1 MW. (For comparison, the 1.76 Linpack petaflops on Jaguar draws close to 7 MW.) The second-generation TSUBAME will also incorporate between 500 TB and 1 PB of SSD storage, and the whole thing is supposed to fit in around 65 racks. They’re scheduling deployment for October 2010.
Assuming they hit their mark, TSUBAME 2.0 will represent more computational power than all the other HPC systems in Japan put together. And since Japan looks like they’re pulling the plug on their Next-Generation Supercomputer Project, which was supposed to field a 10-petaflop machine in a couple of years, TSUBAME may hold that distinction for quite awhile.