The World's Fastest Supercomputer Now in China - PC Magazine

Yesterday's news that China's Tianhe-2 supercomputer is now the fastest in the world isn't all that shocking. After all, an earlier version was on top of the supercomputing list back in 2010. What is a bit more surprising is the architecture behind the Tinahe-2 (also known as Milky Way-2). It is based on Intel's new Xeon Phi architecture, which combines a large number of x86 cores into a single chip; the supercomputer then combines those chips into a single architecture. This system wasn't expected to be deployed for another couple of years, so I was surprised to see a Xeon Phi-based system to top the list. What I find most fascinating here is the competition with systems based mostly on GPU computing. Indeed, a system based on Nvidia's CUDA GPU cores, which topped the list last time, is now in second place. The Top500 list of the world's fastest computers typically comes out twice a year: once in conjunction with the International Supercomputing Conference (ISC) that's happening now in Germany and again at the Supercomputing Conference (SC 13) in the fall. The Tianhe-2, which is based at the National University of Defense Technology in Changsha, China, shows sustained performance of more than 33.8 petaflops (more than 17,500 trillion floating point operations per second) and peak performance of 54.9 petaflops on the LINPACK benchmark. This makes it about twice as fast as the previous leader, the Titan system based at the U.S. Department of Energy's Oak Ridge National Laboratory (ORNL). Tinanhe-2 has 16,000 nodes, each with two Intel Xeon E5-2692 (12-core processors, using the IvyBridge) processors and three Xeon Phi processors for a combined total of 3,120,000 computing cores. The Xeon cores are based on an upcoming 12-core version of the Xeone #5-2600 family, based on the 22nm Ivy Bridge architecture. The total system draws 17.8MW, the most of any top system on the top 500 list, but because the performance numbers are so high, it's still considered relatively power efficient. The June list of the most efficient supercomputers, the Green500, is due out shortly. ORNL's Titan system, which topped the previous list, is now in second place. This is based on a Cray XK7 system with 18,688 nodes, each containing a 16-core AMD Opteron 6274 and an Nvidia Tesla K20x graphics processing unit (GPU) accelerator. This system shows sustained performance of 17.5 petaflops (more than 17,500 trillion floating point operations per second) and peak performance of over 27 petaflops on the LINPACK benchmark. The Sequoia system at the Lawrence Livermore National Laboratory, based on IBM's BlueGene/Q system and its Power CPUs, came in second on the Top500 list a year ago but has slid to third place. The fourth place system remains the "K computer" at Japan's RIKEN Advanced Institute for Computational Science based on Fujitsu SPARC64 processors.   The top four systems show four very different architectures. Traditional big iron systems, such as those based on IBM's BlueGene (Power) and Fujitsu's SPARC architectures, are still very much in the running but most of the attention is going to Intel's new Xeon Phi architecture and Nvidia's CUDA architecture. Meanwhile, there continue to be stories that China is working to create its own processor for supercomputing. In more details, Nvidia announced yesterday that researchers at Stanford University are using GPUs to create the world's largest artificial neural network designed to model how the human brain learns. It also revealed that its CUDA toolkit will now support ARM-based platforms. As part of the supercomputing conference, Intel also introduced new versions of its Xeon Phi coprocessor family, including the 7100 with 61 cores clocked at 1.23GHz, 16GB of memory capacity support, and over 1.2TFlops of double precision performance; the Xeon Phi 3100 family with 57 cores clocked at 1.1GHz and 1TFlops of double precision performance; and a new 5100D, designed so that sockets can attach to a mini-board for use in blade form factors. Intel says the next generation, known as "Knights Landing" and based on the upcoming 14nm process technology, will work not only as a coprocessor, but also as a primary processor, thus removing the complexity of moving data to different pools of memory. This will integrate on-package memory to speed performance. Intel calls the combination of traditional Xeon and Xeon Phi processors "neo-heterogeneous architecture." The hardware architecture has multiple classes of compute capabilities that are accessed by a common programming model. The company stresses that since it's all x86 that could streamline development and optimization in ways that would be harder when using a combination of CPUs and GPU accelerators. Nvidia and the other companies pushing GPU compute would disagree with that assessment.      Intel also talked about using high-performance computing not only for traditional uses such as government and military research and high-end commercial applications like oil and gas simulation, but also for applications such as big data. The goal is to make supercomputing more mainstream. See the full Top500 supercomputers list here. via Science - Google News http://news.google.com/news/url?sa=t&fd=R&usg=AFQjCNEfA4pijEy7rxKjCfIThwqjPiaqsA&url=http://forwardthinking.pcmag.com/computing/312732-the-world-s-fastest-supercomputer-now-in-china

Put the internet to work for you. via Personal Recipe 2954071