More details are emerging about a Chinese supercomputer that some think could become the fastest in the world.
According to HPCwire, the Tianhe-2 is being developed in the National Supercomputer Center in Guangzhou as the successor to the Tianhe-1A, which can drive its 14,336 Intel Xeon X5670 processors and 7,168 nVidia Tesla GPUs to a peak performance of 4.7 petaflops. That performance has put the TH-1 at the top of the list of the world’s fastest supercomputers in 2009; it currently sits at No. 8. (The current champion is Titan, a Cray XK7 system installed at Oak Ridge National Laboratories with a peak performance of 20 petaflops, or 20,000 trillion calculations per second.)
The Tianhe-2 (which has still not formally been unveiled) can reportedly deliver peak performance between 53 petaflops and 55 petaflops, with a LINPACK score for sustained performance of between 27 petaflops and 29 petaflops. (The system was originally slated for delivery in 2015 with a peak performance capability of more than 100 petaflops, a number that’s evidently been significantly revised.) Rumors about an almost-completed version began to leak out following a high-performance computing conference in China last month.
Most of the detail about Tianhe-2 came from a draft of a report (PDF) on the system by Jack Dongarra, a researcher at Oak Ridge National Laboratories and founding member of the Top500 list organization (as well as Distinguished Professor of Electrical Engineering and Computer Science at the University of Tennessee.) According to him, Tianhe-2 runs a total of 3,120,000 processor cores divided among 16,000 nodes. Each node contains sockets for two Intel Ivy Bridge processors that serve as the CPU and three boards for co-processors, which are based on Intel’s Xeon Phi processor, the first commercial incarnation of Intel’s MIC (Many Integrated Core Architecture), which can deliver as much as a teraflop per chipset.
The system is configured with a total of 1.404 petabytes of memory and a parallel storage system with 12.4 petabytes of space. Much of its scalability and performance comes from a proprietary, Chinese-developed network called the TH-Express 2, and a frontend system using 4,096 Galaxy FT-1500 CPUs–16-core processors developed by NUDT based on the Sparc-V9 chip, Dongarra wrote.
Each FT-1500 delivers a total of 144 gigaflops/second, compared to 211 Gflops/s from the Ivy Bridge processors. At peak power consumption, the Tianhe-2 pulls 17.6 megawatts plus another 6.4 megawatts for a closed-coupled water-cooling system, for a total of 24 megawatts. The system runs a NUDT-designed Linux distribution called Kylin Linux, which also runs on the Tianhe-1A.