It’s official: “Titan” is the world’s most powerful supercomputer.
On Monday, Hans Meuer and his team of researchers released the second, semi-annual TOP500 list of the world’s most powerful, general-purpose systems in common use for high-end applications. The 40th edition of the list, released at the SC’12 supercomputer show in Salt Lake City, marks the 20th anniversary of the original TOP500 release.
The list of continually improving, high-performance computers now includes 23 with petaflop class performance, just four and a half years since the debut of Roadrunner, a supercomputer built by IBM for the Department of Energy. Roadrunner has since been pushed to 22nd on the list. The top 5 includes Titan, the DOE’s Sequoia, Japan’s K Computer, the DOE’s Mira, and Germany’s JUQUEEN supercomputer.
When the DOE’s Titan supercomputer was announced at the end of October, the agency and the manufacturer, Cray, claimed that the supercomputer would eventually be capable of over 20 petaflops of theoretical peak computing power, which would place it at the top of the list. That proved to be so, as the submitted theoretical peak (Rpeak) performance of Titan estimated at 27,113 TFLOPS, enough to beat the second supercomputer on the list, Sequoia, by 34.7 percent.
As submitted, Titan includes 560,640 cores, generating a LINPACK (Rmax) score of 17,590, versus 16,325 for the Sequoia. The system combines 2.2-GHz, 16-core AMD Opterons with a Tesla K20 GPU accelerator, connected with a Cray Gemini interconnect. All told, the machine includes 710,144 Gbytes of memory. It also consumes 8209 kW, a 4 percent increase over the Sequoia.
Intel-based systems continue to dominate the supercomputing world, with 380 systems or 76 percent of the TOP500 systems on the list. AMD’s Opteron finished second, with 60 systems or 12 percent—unchanged from six months ago. IBM’s POWER chips captured 53 systems, or 10.6 percent.
Five of the world’s ten fastest supercomputers reside in the United States, including the Titan, Sequoia, Mira, the Stampede supercomputer scheduled to be upgraded at the end of January</a>, and a DARPA trial manufactured by IBM. In all, the U.S. is clearly the leading consumer of HPC systems with 251 of the 500 systems (versus 252 on the June 2012 list, the TOP500 organizers said). The European share (105 systems, down from 106) is lower than the Asian share (123 systems, versus 122). Evaluating the list by number of systems, China ranks second behind the United States. If ranked by aggregate performance, Japan is second, according to Meuer and his team.
From a technical standpoint, several trends are evident, Meuer and his team reported. First, like Titan, another 61 systems on the list (up from 58 during the June list) use some form of accelerator or co-processor. (The “Titan” Cray XK7 system contains 18,688 nodes, with each holding a 16-core AMD Opteron 6274 processor and an NVIDIA Tesla K20 GPU accelerator.) Like Titan, the Tianhe 1A supercomputer from China uses Nvidia Tesla chips, while the Stampede relies on the Intel Xeon Phi as a co-processor.
InfiniBand technology now provides interconnects on 226 systems, up from 209 systems, making it the most-used internal system interconnect technology, the TOP500 organizers said. Gigabit Ethernet interconnects are now found on 188 systems, down from 207 six months ago.