NVIDIA GPUs Now Power World’s 10 Greenest Supercomputers

Written by Edward Chester

November 23, 2013 // 4:04 p.m.

The just-published Green500 list of the world’s most energy-efficient supercomputers shows that the top 10 systems are all powered by NVIDIA Tesla GPUs.

The only other architecture ever to have cleanly swept the 10 top spots on the list is IBM’s legendary BlueGene system.

Crowned the greenest supercomputer, the Tsubame-KFC system at the Tokyo Institute of Technology, hit a record 4.5 gigaflops per watt. That’s about 25 percent more efficient than the list’s number-two, Cambridge University’s Wilkes, at 3.6 gigaflops per watt. In third place was the system at Japan’s Center for Computational Sciences, at the University of Tsukuba, at 3.5 gigaflops per watt.

NVIDIA’s presence at the top of the Green500 is in the ascendant. Just six months ago, only two of the top 10 systems were accelerated by our GPUs.

At the heart of this trend is the spread of NVIDIA Tesla GPU accelerators based on our Kepler architecture. Launched last year, they are three times more energy efficient than the Fermi-based family of processors they succeeded.

Efficiency has, obviously, become an increasingly key consideration for supercomputing systems as they’ve gotten faster. The largest supercomputers can consume megawatts of power, pushing their annual energy costs into the tens of millions of dollars.

Improving energy efficiency is central to achieving exascale computing – that is, delivering supercomputers that run 50 times faster than today’s best – at a speed of 1 exaflops, or a million trillion flops.

For that, we’ll need a tenfold increase in efficiency – with systems running at 50 gigaflops per watt, according to an SC13 talk by the University of Tennessee’s Jack Dongarra, at NVIDIA’s GPU Technology Theater today.

That may, on its face, sound challenging. But bear in mind that the new Green500 list shows that we’ve achieved a better than tenfold gain in efficiency since the November 2007 list, when the top system, at England’s Daresbury Laboratory, was running at only 0.3 gigaflops a watt.