How GE Uses GPUs To Chase Wind Energy in Chilly Climates

by Brian Caulfield

Wind power could make up as much as 5 percent of the global electricity supply within a decade, nearly twice its current level.  The problem: many of the world’s windiest spots are in cold climates where ice slashes the efficiency of wind turbines.

That’s why General Electric, one of the world’s largest wind turbine manufacturers, is using the Tesla-powered Cray XK7 Titan supercomputer at the Oak Ridge National Laboratory to simulate how hundreds of millions of water molecules freeze on the surfaces of those huge, whirling blades.

Speaking at the Supercomputing 2013 summit in Denver, from NVIDIA’s GPU Technology Theater, Masako Yamada, a physicist at the GE Global Research Advanced Computing Lab described how Titan’s GPUs made it possible for her team to model vast numbers of water molecules over a wide variety of surfaces.

Putting wind turbines in colder climates could help push the portion of the global electricity supply generated by this renewable resource up to 5 percent.

To deal with freezing conditions, today’s turbines rely on heaters in the blades to melt ice. But these can consume up to 10 percent of the energy the turbine produces. The solution: creating surfaces that make it tough for ice to form in the first place.

That’s no easy proposition. For one, different materials work to stop ice from forming at different temperatures.

Finding the right material for GE’s turbines meant Yamada’s team would have to understand when ice begins to form – down to the quadrillionth of a second – and understand how the process works on an atomic level for hundreds of droplets, consisting of millions of molecules each.

Titan’s K20 accelerator’s helped Yamada’s team speed up their work.

That’s a huge computing problem. Solving it requiredrunning simulations of unprecedented scale. So Yamada turned to Oak Ridge’s Jaguar supercomputer, and then to its successor, Titan. Both pair CPUs with GPUs to wring more work out of each watts of power.

To get the most out of these machines, Yamada’s team worked with Oak Ridge computational scientist Mike Brown to repurpose the LAMMPS molecular dynamics application to take advantage of Titan’s NVIDIA K20 Tesla accelerators, which can race through repetitive tasks much more efficiently than CPUs.

“We were able to achieve 5x acceleration compared to half a year ago by porting our CPU code to hybrid CPU-GPU code,” Yamada said, making her million molecule droplet models possible.

Guess you could say GPUs knocked her problem out, um, cold.