How One Researcher Used 18,000 GPUs to Simulate the Most Glorious Event in Astrophysics

by Brian Caulfield

It’s one of the most glorious events in all of astrophysics.

Streams of particles — so long they can only be measured in light years — shower out from stars, black holes and even galaxies. It’s a phenomenon that’s been as mysterious as it is lovely, thanks to the vast distances between Earth and these amazing structures. Until now.

The Tesla-powered Titan supercomputer helped researchers better understand one of the universe’s most spectacular sights.

Speaking at the Supercomputing 2013 summit in Denver, from NVIDIA’s GPU Technology Theater, German researcher Michael Bussman spoke about how his team used the NVIDIA Tesla-powered Titan supercomputer at the Oak Ridge National Laboratory to better understand one of the night sky’s most amazing spectacles.

As a result of its work, Bussman and his team, who work at the Helmholtz-Zentrum Dresden-Rossen research laboratory, are finalists for the 2013 Gordon Bell Prize, the highest award in the field of supercomputing.

Their team started by modeling the Kelvin-Helholtz Instability, or KHI, which occurs when plasma jets collide. KHI simulations are done in much the same way as simulations performed for fluids. That’s because the electromagnetic fields inside plasma streams act a lot like the currents inside streams of fluids.

The problem: the team had to account for the behavior of individual particles, which create fields that, in turn, change the behavior of the particles. That meant interactions between particles had to be modeled as well.

That’s where Titan comes in: Bussman’s team was able to harness roughly 18,000 of the machine’s GPUs to model streams of unmagnetized hydrogen plasma — comprised of 75 billion particles — and work from there to model the radiation streaming from those jets.

The result is the biggest, highest-resolution kinetic KHI simulation to date, according to Bussman. Credit goes to Titan’s GPUs, which can tear through the repetitive calculations needed to build such simulations faster than any CPU.

Now, scientists can look at the radiation seen from these jets and compare them to Bussman’s simulations. When they find a pattern that matches one of Bussman’s models, they’ll better understand the behavior of particles that generate the radiation captured by their Earth-bound instruments.

Video streaming by Ustream