Russian Supercomputer Adopts Xeon Phi Share your comment!


The Russian Academy of Science is aiming for the region’s fastest supercomputer in a joint effort with the Commonwealth of Independent States (CIS, formerly the Soviet Republics). By harnessing thousands of parallel Intel Xeon and Xeon Phi cores, the 10 petaFLOPS JSCC supercomputer will advance the frontiers of Soviet science. 

The Joint Super Computer Center (JSCC, Moscow) provides high-performance computing (HPC) services to over 100 scientific and educational communities as well as to 181 independent research labs working on over projects in Russian and the CIS.

By paring Intel Xeon E5-2690 processors with Xeon Phi coprocessors in the RSC Group’s Tornado parallel architecture using direct liquid cooling, JSCC claims to have tripled the performance of its supercomputer center while reducing its total cost of ownership (TCO) by 25 percent. Now in the prototype stage, the new supercomputer aims to set new records in energy efficiency, which it estimates will exceed 1,949 megaFLOPS per watt–over five times better than the most energy efficient supercomputer in Europe, representing a 60 percent energy cost saving.

In all, the JSCC supercomputer will house 208 Intel S2600JF server boards—each with two 2.9-GHz Xeon E5-2690 processors that each have eight cores–housing a total of 3,328 parallel Xeon E5-2690 cores. Each server board will also connect by PCIe to two Xeon Phi coprocessor cards–416 total–each with 61 cores for a total of 25,376 Xeon Phi cores. Over 24 terabytes of mass storage uses Intel’s solid-state drives (SSDs) connected to the parallel processors with Intel’s fourteen data rate (FDR) Infiniband optical fabric. When finished, the JSCC supercomputer will top 10 petaFLOPS peak performance, however the prototype is currently running at just under 524 teraFLOPS (over 375 teraFLOPS when running the LINPACK benchmark). And with no fans or other moving parts–due to the absence of traditional hard disks in favor of its SSD arrays–the new supercomputer is virtually noiseless, consuming just 222-kWatts in its current 524 teraFLOP configuration.

Scientists at JSCC report that the same parallel software development tools and programs they already use have been capable of easily porting its existing x86 applications over for execution on the new supercomputer accelerated by its massive array of Xeon Phi cores. And by using the liquid cooled RSC Tornado architecture, the footprint of the new supercomputer has achieved the industry’s highest computing density of 141 teraFLOPS per cubic meter (or about 181 teraFLOPS per rack)–what JSCC claims is 3.8-times denser than the previous world record for x86 supercomputers.

Posted on by R. Colin Johnson, Geeknet Contributing Editor