PORTLAND, Ore. — For the third time in a row it was not even close: China’s Tianhe-2 (named after the Milky Way) at the National Super Computer Center in Guangzhou had more than twice the peak performance (54.9 petaFLOPS) than the second-place winner (27.1 petaFLOPS). And no wonder — Tianhe-2 houses more than five times as many Intel Xeon and Xeon Phi cores (3,120,000) than the second-place Titan at Oak Ridge National Laboratories — 560,640 AMD Opteron cores with Nvidia graphics processing units (GPUs). In fact, the only change since the last Top500 ranking in November 2013 was the tenth place, which was taken over by a secret-location US government Cray XC30, with 225,984 Intel Xeon cores peaking out at 4.9 petaFLOPS.
When running the Linpack benchmark, the Intel Xeon powered Tianhe-2 scored just shy of twice (1.93-fold) the maximum benchmark performance of the second-place AMD Opteron powered Titan — 33.9 petaFLOPS compared to 17.6 petaFLOPS. In fact, Intel Xeon and Xeon Phi cores powered almost 86% of the Top500 entries, IBM Power processors was next at 8% and AMD Opteron processors were used in 6% of the Top500.
Coprocessor architectures were on the rise too, with a total of 62 systems using coprocessors, up from 53 in the November 2013 listing six months ago. Nvidia GPUs powered 44 systems, 17 systems were powered by Intel Xeon Phi coprocessors, and 2 used ATI’s Radeon GPUs. In the Top10, first-place Tianhe-2 and seventh-place Stampede used Intel Xeon Phi coprocessors. Second-place Titan and sixth-place Piz Daint used Nividia GPUs.
The US retained its top position with 233 systems in the Top500, but this was down from 265 on the November 2013 list. Chinese systems ranked next with 76, up from 63 last November. The Japanese also rose to 30, up from 28. The UK was next with 30, France with 27, and Germany with 23.
As far as system builders go, Hewlett Packard led with 197, or 39% of the systems; IBM was next with 164 systems, or 33%; and Cray came in third with 48 systems, or 10%. Also, 37 of the Top500 systems had greater than one petaFLOPS of performance, up from 31% in November 2013. The total number of petaFLOPS for all 500 systems was 274, compared to 250 six months ago.
Click the Tianhe-2 image to see a slideshow of all top 10.
— R. Colin Johnson, Advanced Technology Editor, EE Times