Fast Radio Bursts (FRBs) are extremely bright flashes of radio light, that travel billions of light years to reach Earth. Discovered over a decade ago, their origin and nature are still largely a mystery. Because the flashes last only a fraction of a second, they are easy to miss and very difficult to observe. Therefore, only about 25 FRBs have been discovered so far.
That is now going to change with Apertif, the new wide-field cameras for ASTRON's radio telescope in Westerbork, the Netherlands. Apertif has the largest, most sensitive field of view of all radio telescopes in the world. To find FRBs, Apertif needs to continuously make a high-speed movie of the radio sky, at 20.000 frames per second. This requires new, more powerful brains. "To form and process all those images, we need the computing power that only the fastest supercomputers in the world can produce", stated Joeri van Leeuwen from ASTRON and the University of Amsterdam. "But we did not have such a computer yet. That's why we designed and built this one ourselves."
ClusterVision, Europe's dedicated specialist in high performance computing solutions, has completed this new high performance computing GPU cluster system for the Netherlands Institute for Radio Astronomy (ASTRON). The 2 PFLOPS installation, codenamed ARTS, will be used to assist the institute's Westerbork Synthesis Radio Telescope with analysing and deciphering the large pulsar flashes.
The institute was looking for a state of the art HPC solution that would satisfy all their needs. ClusterVision designed a GPU based cluster that is able to process all this data. By employing a large number of GPU nodes, data from the telescopes could be processed much faster and way more precisely.
The new supercomputer has a special feature. It is completely powered by image processing chips from the gaming industry. "Gamers use very powerful processors for video tasks: the GPUs", explained Joeri van Leeuwen. "We now use these chips for the first time to process the high-speed images of our telescope." The supercomputer consists of 200 of these GPUs, which process 4 terabits of data per second: more than the entire internet of the Netherlands. With a peak compute capacity of 2 petaflops it is the most powerful GPU supercomputer in the Netherlands.
Furthermore, by utilising the deep learning capabilities a GPU cluster brings to the table, the telescopes will be able to detect pulsar flashes with much greater accuracy through self-learning. In the past, ASTRON scientists had to manually detect and input pulsar patterns. With deep learning, ARTS does it for them.
The supercomputer will be able to teach itself to find FRBs in the thousands of pictures of the telescope. "We have always done this manually", stated Joeri van Leeuwen. "But that is a lot of work, and also error-prone. As it detects more flashes, the supercomputer will learn to distinguish the flashes better and better. We hope to discover one FRB per week. Our system is unique in the world in that it can then also immediately determine the precise location."
With the supercomputer Joeri van Leeuwen hopes to unravel the mystery surrounding the Fast Radio Bursts. "We know that they come from other galaxies, but we cannot pinpoint the exact location. It is also unknown whether all FRBs are bright bursts from neutron stars, exploding stars, or black holes that send out flashes." Astronomer Samayra Straal from the University of Amsterdam is also eager for the output of the new machine. "Some FRBs repeat. With the new technology, we discover the flash 'live', and can immediately freeze the telescope motion to stare in that direction for more flashes. We think that can help us understand what causes them."
The Apertif supercomputer, called ARTS, was funded by the Netherlands Institute for Radio Astronomy (ASTRON), the Netherlands Organisation for Scientific Research (NWO) and the Netherlands Research School for Astronomy (NOVA).