Canada's Condominium Magazine
Upon learning that a pair of NVIDIA-powered supercomputers had taken the top two spots on the Green500 list, published today, making them the two most energy-efficient supercomputers in the world, we wondered what that meant.
The two supercomputers, it turns out, are powered by NVIDIA Tesla K20 GPU accelerators, and the first-place winner, called Eurora, delivers 3,210 MFlops per watt. It beat the Intel competitors by a mile, being twice as energy-efficient as the Intel Xeon Phi system, known as Beacon.
But again, what does that mean? Should we all have one of these energy-efficient supercomputers on our desks?
Well, not really. Supercomputers are actually collections of thousands of “nodes” acting in parallel. Their operating speed is measured in Flops, which stands for Floating-point Operations Per Second. The world’s most powerful supercomputer, just announced a few days ago, is a Chinese monster. Its speed is given as 33.86 petaFlops. A petaFlop equals 1 quadrillion computations per second, which looks like this: 1,000,000,000,000,000 (one thousand trillion). The computational speed of your “average” desktop with a 2.0 GHz processor is measured in the billions.
It is unlikely, however, that the Chinese supercomputer will hold on to its title for long, as computer developers everywhere are constantly pushing the limits of these things. The new goal is to develop exascale computing by the year 2018. An exascale computer will be capable of at least one exaFlop, which is a one thousandfold increase over a petaFlop, or one quintillion computations (1,000,000,000,000,000,000).
Powering all that computation is expensive. They require “egregious” amounts of power and they generate so much heat that they must be cooled by “extravagant” cooling facilities. The world’s current number three supercomputer, called Jaguar, which stretches across an entire rather wide room, uses 7 megawatts (MW) of power, costing more than $1 million a year just to keep it turned on. But an exascale computer being developed by the US Defence Advanced Research Projects Agency (DARPA) will use as much as 400 MW of power, at a cost of about $300 million a year.
Which is why this energy-efficient supercomputing competition is so important. Using the Eurora model—that’s this year’s winner, the current most energy-efficient supercomputer in the world—the “power envelope” of DARPA’s exascale machine could be reduced by 24 per cent, dropping it from 400 MW to 312 MW. This, the Green500 people note drily, is “a step in the right direction.”
The Green500 hopes to raise awareness among the supercomputing community that there is more to computing than just speed. There is also “energy efficiency, reliability, availability and usability.”
And what do they do with these supercomputers? Well, as the DARPA reference indicates, they are used in defence-related work, simulating the effects of a nuclear explosion, for example. They also play a big role in climate modelling, weather forecasting, seismology, astrophysics and in the field of molecular dynamics, which studies the behaviour of atoms.