Water-Cooled Parallel Processors Use MIC Share your comment!


Water-cooled parallel processors based on Intel’s Many Integrated Core (MIC) architecture are boosting performance and cutting energy consumption in Russian research centers. 

Common to all the Russian parallel processors, including that of the previously announced Russian Academy of Sciences, is direct-liquid cooling of their Xeon and Xeon Phi cores using the RSC Tornado architecture, which pipes coolant to the parallel processors where most of the heat is generated, removing up to 100-kiloWatts of power from standard 42U racks housing up to 128 server boards.

 The unique RSC direct-liquid cooling design uses “hot water” of up to 130 degrees Fahrenheit (55 degrees Celsius) for improved heat exchange–up to 4000-times more efficient than direct air cooling. The result is a power density of more than 25 teraFLOPS per square meter at a ultra-green Power Usage Effectiveness (PUE) of 1.2.

The systems include a major 237 teraFLOPS supercomputer at the South Ural State University (SUSU) as well as a smaller 83 teraFLOPS supercomputer at the Moscow Institute of Physics and Technology (MIPT) and a 35 teraFLOPS supercomputer at the Russian Federal Service for Hydrometerorology and Environmental Monitoring.

The new Xeon Phi powered supercomputer at the stately South Ural State University (SUSU, pictured) serves the research interest of the largest student population in Russia. SOURCE: Intel

The SUSU installation will assist researchers with a wide range of parallel applications addressing both theoretical and practical scientific and engineering disciplines–54 percent for engineering modeling, 42 percent for natural sciences, and four percent for social and economic forecasting. More than 250 separate development projects are already underway by SUSU researchers where the new supercomputer enables detailed models and simulations of metallurgy, mechanical engineering, energy and computer manufacturing techniques to accelerate development in the region.

“Computer models and virtual test beds are being developed in order to conduct computing experiments to find effective solutions. Then, when all the supercomputer calculations are complete, an actual solutions can be created based on our models,” said Leonid Sokolinsky, dean and director of the Supercomputing Simulation Laboratory at SUSU.

The SUSU supercomputer uses 192 server blades with a total of 384 Xeon and 192 Xeon Phi coprocessors for a total of 14,804 x86 cores occupying just 538 square feet (50 square meters) of floor space. Each direct-liquid cooled server uses two octal-core Xeon E5-2690 processors whose performance has been boosted with an attached 61-core Xeon Phi coprocessor, resulting in a 60 percent reduction in energy costs or about 995 megaFLOPS per watt.  The SUSU supercomputer is also using Intel’s solid-state drives (SSDs) instead of hard disks, interconnected with quad data rate (QDR) Infiniband fabric. Besides a peak performance of over 236 teraFLOPS, the SUSU supercomputer delivers over 146 teraFLOPS when running the LINPACK benchmark.

Posted on January 29, 2013 by R. Colin Johnson, Geeknet Contributing Editor