Peering Into the Sky Deeper and More Sharply with GPUs

by Tony Kontzer

These are great days for astronomers, who have access to increasingly powerful telescopes that generate unprecedented amounts of data. But with all that data come challenges.

In particular, a new generation of software-based telescopes—most notably the huge Low Frequency Array, or LOFAR, built by the Netherlands Institute for Radio Astronomy—are serving notice that more powerful processing capabilities are needed, said John Romein, senior researcher at the institute, speaking at the GPU Technology Conference.

“We want to build GPU technology into the current generation of telescopes like LOFAR,” Romein said.

LOFAR owes its heavy data load to one of its primary innovations – its omni-directional antennas which represent a breakthrough in sensitivity. More sensitivity means more data—245 Gb per second, to be precise—which in turn means new scientific opportunities.

And things figure to get even more complex with the construction of the so-called Square Kilometer Array (SKA), a radio telescope being built in South Africa and Australia that Romein said will be bigger than all of the world’s current radio telescopes combined. Once finished, the SKA will produce a volume of data magnitudes of order greater than LOFAR.

With so many telescopic advances pushing the limits of technology, Romein and his team are looking into how various GPU accelerators can help with processing the massive flow of data.

“We want to get an understanding of what properties make an accelerator efficient,” Romein said.

What the ASTRON team has found is that NVIDIA GPUs compare favorably with the likes of Intel’s Xeon Phi and AMD’s FirePro dual GPU 3D graphics card. For instance, while the FirePro S10000 tested faster than the Tesla K10, it also drew significantly more energy.

The team still plans to test a number of other considerations, including more detailed analysis of energy efficiency, relative programmability, any optimizations required to implement signal processing algorithms, and the degree to which each accelerator is dependent on the architecture.

“There’s quite a bit of work to do to get the code ready for use in LOFAR,” said Romein.