Twice a year, the TOP500 ranking of supercomputers shows which machine can claim bragging rights as the world’s most powerful, often a quickly shifting title in the fast-moving world of supercomputing. But the TOP500 list released publicly today is astounding for how little its upper rankings have changed since June. Indeed, the computers in the top 11 positions have remained exactly the same.
That’s not to say that there’s been no advancement on the supercomputing front. The highest-ranking supercomputer, the K Computer at Japan’s RIKEN Advanced Institute for Computational Science in Kobe, was able to perform 8.2 petaflops (that’s 8.2 x 1015 floating-point operations per second) back in June, whereas in the ranking released today it broke the 10-petaflops barrier at 10.5 petaflops.
When June’s TOP500 ranking appeared, we noted that the K Computer had nudged out China’s Tianhe-1A, the previous record holder, and we forecast that the K Computer would maintain the title for another year or two perhaps. So it’s satisfying to see that it has indeed kept the number-one position in the list announced today. Of course, at Spectrum we like to look beyond the mere ranking of speed and search for larger trends in high-performance computing.
Last June, for example, we commented on the rate at which supercomputing advances. And in November 2010, we spoke with University of Tennessee professor Jack Dongarra, one of the computer scientists behind the TOP500 list, who pointed out that both the number-one and number-three machines at the time (both of which were Chinese) have heterogeneous architectures that combine multicore CPUs with GPUs to achieve their impressive computing might. So the take-away message last year certainly seemed to be that heterogeneous designs were the ticket to high-power, energy-efficient computing.
This year’s rankings do, however, call that conclusion into question. The K Computer contains 68,544 eight-core SPARC64 VIIIfx CPUs with no GPUs. Yet the machine is nevertheless plenty energy efficient, being able to perform 830 megaflops per watt of power expended. For comparison, the Tianhe-1A (which now ranks number 2 in computation speed) provides just 635 magaflops per watt. The fourth-ranking Dawning TC3600 system, also built in China and also using a combination of CPUs and GPUs, looks even worse when judged from the standpoint of energy efficiency: 493 megaflops per watt. The most efficient machine on the list, the Blue Gene/Q prototype at IBM’s Thomas J. Watson Research Center in Yorktown Heights, New York, which uses just multicore CPUs, can perform more than 2000 megaflops per watt.
So we were wrong a year ago to imply that heterogeneous architectures were the way of the future? Maybe not. While it doesn’t rival the Blue Gene/Q prototype, the most energy-efficient supercomputer of the 10 fastest supercomputers in today’s ranking is number 5 in terms of overall computational speed: the Tsubame 2.0 system at the Tokyo Institute of Technology’s Global Scientific Information and Computing Center, and that system uses both CPUs and GPUs.
Indeed, a ranking the worlds top supercomputers by energy efficiency gives a mixed picture, with some of the most energy-efficient machines using GPUs in addition to CPUs and others not. Given the importance of energy efficiency for the future of supercomputing, it would seem that the ideal architecture for building super-powerful computing machines is still up for grabs.