SAN JOSE, Calif. – Intel’s Xeon Phi co-processor made a strong showing on the latest list of the world’s Top 500 supercomputers, powering seven of the systems including one in the top 10. Nvidia’s graphics chips continued to dominate the coprocessor category, and the Infiniband interconnect widened its lead over Ethernet as the top interconnect.
Titan, a Cray XK7 system installed at Oak Ridge National Lab, ranked as the world’s most powerful system at 17.59 petaflops on the Linpack benchmark. It uses nearly 300,000 AMD Opteron processors and 261,632 Nvidia K20x graphics chips as accelerators.
Although it consumes a whopping 8.21 megawatts, the Cray Titan edged out IBM’s BlueGene/Q architecture as more energy efficient, delivering 2,143 Mflops/W. BlueGene/Q was the most popular architecture in the top 10, used by the second, fourth, fifth and ninth most powerful systems.
Twenty-three systems on the latest list cracked the petaflops milestone. Roadrunner, the first petaflops-class system, debuted on the Top 500 four-and-a-half years ago.
The use of massively multicore coprocessors is fueling the rapid rise in performance. Sixty-two systems on the latest Top 500 list used coprocessors, up from 58 six months ago. Fifty of the supercomputers on the latest list use Nvidia graphics chips as coprocessors.
Intel’s Xeon Phi appeared in seven of the systems including Stampede, a Dell PowerEdge C8220 system at the Texas Advanced Computing Center in Austin, Texas. Stampede hit 2.6 petaflops, good for seventh place on the list.
Xeon Phi surpassed accelerators such as the AMD Radeon and IBM Cell, used in three and two systems, respectively.
“There’s a need for energy efficient architectures,” said Erich Strohmaier of Lawrence Berkeley National Laboratory, one of four researchers who compile the Top 500 list. “IBM's BlueGene/Q architecture as well as accelerators from Nvidia, Intel and others offer different alternative paths towards this goal, and currently there is no clear winner among them,” he said.