Japan’s K Computer surprised the world with a giant leap in floating point number encoding performance today and is approaching the estimated calculation capability of the human brain.
Japan’s Riken Advanced Institute for Computational Science has taken the first spot on the prestigious list of the world’s 500 fastest supercomputers. The K Computer sets a new record in the number of processing cores, power consumption as well as peak and sustained computational performance. The current installation is not yet completed, but is already rated a peak performance of 8.77 PFlop/s, which equals 8.77 quadrillion floating point operations per second (Flop/s). The maximum sustained performance is 8.2 PFlop/s, more than three times the sustained performance of Tianhe-1.
K Computer
Other than Tianhe-1, the Chinese supercomputer that claimed the first position in November of last year with a peak performance of 4.7 PFlop/s, K does not use GPUs to increase its theoretical peak floating point performance. The use of GPUs has been viewed with a critical eye lately and has even been considered as cheating in an effort to achieve a high Top500 ranking. It’s extremely difficult to exploit the full the theoretical potential of a GPU supercomputer.
K integrates 68,544 2.0 GHz Fujitsu Sparc VIIIfx CPUs with 8 cores each, which makes K the first supercomputer with more than half a million cores (548,352). Each core is rated at a peak performance of 16 (double-precision, DP) GFlop/s, which makes the Sparc VIIIfx CPU slightly more capable than Intel’s latest Sandy Bridge CPUs. A quad-core Sandy Bridge chip is estimated to deliver about 100 DP GFlop/s. The Fujitsu processor is not a mainstream processor, but was developed specifically for supercomputing in mind. The 45 nm chip features are dedicated floating point instructions that enable the processor to run 8 floating point operations per clock cycle per core.
Given its performance, the processor is relatively modest as far as its power consumption is concerned: The 8-core processor is said to only consume 58 watts. That may give K a high power-efficiency ratio, but the absolute power consumed is extremely high: The system is currently running at 9.9 MWatts and incurs an electric bill of about $10 million per year. When finalized, K will hit a performance of about 10 PFlop/s and a power consumption of close to 11 MWatts.
K is reminiscent of the Earth Simulator, which was announced at Japan’s Earth Simulator Center in June of 2002. Back then, the system posted a performance of 35.9 TFlop/s, which was far ahead of the second-fastest system, Lawrence Livermore’s ASCI White, which was rated at 7.2 TFlop/s. K is similarly advanced and it may take some time until other nations catch up. The United States still stands at #3 with Jaguar, a 2.3 PFlop/s supercomputer (1.8 PFlop/s sustained) that is installed at the Oak Ridge National Laboratory. IBM recently said that it is constructing the 750,000 core Mira supercomputer for the Argonne National Library, which targets a peak performance of about 10 PFlop/s, but it is unclear when this computer will be completed. The Earth Simulator held its #1 rank for more than 2 years until November 2011, when it was replaced by two U.S. supercomputers.
Today, the original Earth Simulator would not make the Top500 list anymore. The minimum required performance for the list is currently 70 TFlop/s.
The progress in supercomputing is truly dazzling. The cost to theoretically achieve a performance of 1 TFlop/s was about $1.1 quadrillion in the early 1960s and was first achieved in 1997 with ASCI Red, a Pentium Pro 200 MHz system that was installed at Sandia National Laboratories. Today, 1 TFlop/s can be achieved with an investment of about $2000 with CPUs alone (or substantially less if you are using GPUs). Exascale supercomputers, systems that will exceed a performance of 1000 PFlop/s, are expected to arrive in the 2020 timeframe. Such systems may integrate “hundreds of millions of processing cores”, IBM said in a statement when it announced its Mira supercomputer.
Engineer and inventor Ray Kurzweil believes that the human Brain is capable of a comparable processing power of about 10 PFlop/s. Of course, there are different opinions and a rather scientific experiment to estimate the computational ability of the human mind by IBM engineers concluded that our brain may be closer to about 38 PFlop/s and a memory of about 3584 TB. At the current speed of progress, we may be able to simulate a human brain as early as next year or, by IBM’s estimates, within 3 years. Needless to say, there are supercomputers that easily trump the human brain in specific tasks – remember IBM’s Watson competing in Jeopardy? However, they cannot simulate the complexity of the human brain and its abilities to multitask, especially control our organs.
You can leave a response, or trackback from your own site.
