Slashdot Mirror
Swiss Supercomputer Edges US Out of Top Spot (bbc.com)
There have only been two times in the last 24 years where the U.S. has been edged out of the top spot of the world's most powerful supercomputers. Now is one of those times. "An upgrade to a Swiss supercomputer has bumped the U.S. Department of Energy's Cray XK7 to number four on the list rating these machines," reports the BBC. "The only other time the U.S. fell out of the top three was in 1996." The top two slots are occupied by Chinese supercomputers. From the report. The U.S. machine has been supplanted by Switzerland's Piz Daint system, which is installed at the country's national supercomputer center. The upgrade boosted its performance from 9.8 petaflops to 19.6. The machine is named after a peak in the Grison region of Switzerland. One petaflop is equal to one thousand trillion operations per second. A "flop" (floating point operation) can be thought of as a step in a calculation. The performance improvement meant it surpassed the 17.6 petaflop capacity of the DoE machine, located at the Oak Ridge National Laboratory in Tennessee. The U.S. is well represented lower down in the list, as currently half of all the machines in the top 10 of the list are based in North America. And the Oak Ridge National Laboratory looks set to return to the top three later this year, when its Summit supercomputer comes online. This is expected to have a peak performance of more than 100 petaflops.
Something tells me NSA or other state entities have faster computing power
Why would they? The NSA does cryptanalysis, text scanning, and signal analysis. None of these require a supercomputer.
If your tasks can be run in parallel on distributed systems, then a supercomputer is a waste of money. Supercomputers are for tasks that have Amdahl Bottlenecks, or tight data dependencies.
The system configuration is outlined here: http://www.cscs.ch/publications/highlights/2017/piz_daint_one_of_the_most_powerful_supercomputers_in_the_world/index.html. The storage is listed as Cray Sonexion, which is Lustre-based. The Cray Aries network has also shown itself to scale pretty well. Of course, the LINPACK benchmark that's the basis of the Top 500 list doesn't really measure filesystem or network performance, so the relative merit of those subsystems can't be judged effectively by those results.
The Top500 data also included the results of the HPCG benchmark (http://www.hpcg-benchmark.org/), which has this stated purpose: "HPCG is designed to exercise computational and data access patterns that more closely match a different and broad set of important applications, and to give incentive to computer system designers to invest in capabilities that will have impact on the collective performance of these applications."
The Swiss system is ranked #3 on the HPCG list as well and also #6 on the Green 500 list of most energy-efficient supercomputers, so it's a fairly well-balanced system and not just some machoFLOPS stunt.