The latest figures from the TOP500, a list of world-leading supercomputers, are in - and there are some major wins for the world of GPGPU computing.
It's no surprise that China's Tianhe-1A supercomputer, which we looked at last month, is at the top of the leaderboard with a confirmed Linpack benchmark score of 2.57 petaflops, knocking previous leader, the US supercomputer Jaguar, down to second place with 1.75 petaflops. What may surprise, however, is the entry of a second Chinese system at number three in the list - bracketing the once-world-leading American entry.
The two machines, along with others that don't quite make it into the top five, put China at the number two spot as world supercomputer power, behind the US but ahead of the UK, France, Germany, and Japan - despite the French CEA installing a new supercomputer from Bull which comes in at number six.
Both Tianhe-1A, at number one, and Nebulae, at number three, use Nvidia Tesla graphics processing units alongside Intel Xeon processors to blast their way to the top of the list. The rise of general purpose GPU (GPGPU) computing with Nvidia chips, usually using the company's CUDA language or the open-source OpenCL alternative, is undeniable: three of the top five computers use Nvidia GPUs, and 28 of the 500 in total are using GPGPU computing to achieve their massive performance.
Despite growing interest in Nvidia's offerings, Intel has remained a winner in the TOP500 list. With its Xeon processors often being chosen to partner with Nvidia's GPUs, Intel's chips account for 79.6 per cent of all systems in the list. The company's Westmere line has proven popular, with 56 machines now using the chips compared to just seven in the last update.
While supercomputers often have massive quantities of physical processors - Tianhe-1 has over 14,000 Intel Xeon processors despite concentrating mainly on GPGPU computing - many designers are chosing high core counts to help boost performance. Quad-core chips make up 73 per cent of the systems in the list, while 19 per cent use CPUs featuring six or more cores.
The move to GPGPU computing has brought another benefit: a reduction in the power required. Compared to the same list six months ago, average power efficiency across the TOP500 list has risen to 219 megaflops per watt from 195 megaflops per watt, while only 25 machines on the list draw more than one megawatt of power.
Sadly, that leaves the Oak Ridge National Laboratory Jaguar system behind the curve: the number two machine draws a massive 6.95MW, more than double that of the average across the top ten supercomputers.
If you're interested in building a supercomputer and getting your name in lights, there's some bad news: the influx of high performance machines into the latest list means that the entry level to get your computer listed is now 31.1 teraflops, up from 24.7 teraflops six months ago.
The full list is available on the TOP500 website.