Tianhe-2 notches up 30.65 petaflops during benchmark test
China scaled the supercomputing summit in late 2010, when a 2.507-petaflop machine named Tianhe-1A (or Milkyway-1) was ranked as the fastest supercomputer by TOP500, which publishes a list of the fastest 500 supercomputers twice a year. Its stay there, though, was brief, lasting all of six months. But if you think it was just a flash in the pan, the Chinese National University of Defense Technology (NUDT), the organization which developed the Tianhe-1A, is building the Tianhe-2 in order to prove you wrong.
University of Tennessee professor Jack Dongarra, who is one of the three people responsible for compiling the TOP500 list, visited the Chagsha, China-based NUDT last month for an International HPC Forum (IHPCF), where participants were briefed about the Tianhe-2 supercomputer and even allowed to feast their eyes on the 3,120,000-core beast.
Dongarra was given test results of a Linpack benchmark run that used 14,336 (theoretical peak: 49.19 petaflops) of the TH-2’s total 16,000 nodes (theoretical peak: 54.9 petaflops) and lasted five hours. Despite running at 90 percent capacity, the TH-2 managed a stellar 30.65 petaflops. This is considerably higher than the 17.6 petaflops achieved by the ORNL Titan, currently the world’s fastest supercomputer.
Now, for all you high-performance computing (HPC) aficionados out there, here’s a rundown of the Tianhe-2’s specs:
Dongarra expects the TH-2’s imminent reign at the top to be fairly long, with a response from the U.S. unlikely before 2015. If a response does come before that, it is more likely to come from Japan than the States. This, as Dongarra sees it, is a wake-up call of sorts for the U.S.
You can find Dongarra's detailed report on the TH-2 at this link (pdf).