Seventeen quadrillion operations per second. Twenty million solder joints. Five thousand meters above sea level. Forty-eight months to install.
These are just some of the big numbers behind a newly installed supercomputer designed to run one of the world's most complex ground telescopes, nearing completion on a remote, high plateau in Chile. Among other challenges, engineers had to overcome the site's harsh conditions and low oxygen levels to build the Atacama Large Millimeter/submillimeter Array (ALMA) Correlator, which will ultimately combine deep space signals captured by the telescope's dozens of antennas. "The technical challenges were enormous, and our team pulled it off," said Mark McKinnon, the North American ALMA Project Director at the U.S. National Radio Astronomy Observatory (NRAO) in Charlottesville, Virginia, in a statement.
ALMA, which is expected to be completed in March and is already operating on a limited basis, will use a giant array of 66 antennas to gather electromagnetic signals produced in deep space. To produce useful images and information, however, astronomers must first use a supercomputer—the Correlator—to process the signals and make the array function as a single telescope. But building a supercomputer that can handle signals from 2016 possible antenna pair combinations and operate at 5000 meters above sea level, where oxygen concentrations can be roughly one-half of those found at sea level, is no easy task, Correlator subsystem manager Alain Baudry of the Laboratory of Astrophysics at the University of Bordeaux in France tells ScienceInsider. That is because the thin air makes it harder to cool components, and data storage drives don't work reliably.
To get around those problems, Correlator engineers carefully optimized the design and placement of resistors, capacitors, and other components to reduce energy use and ease air cooling. And data from the diskless Correlator is transmitted down to an ALMA support facility located at 2900 meters, "where it is stored on special disks," Baudry says. Ultimately, copies of the data are transmitted to ALMA partners in the United States, Europe, and Asia.
The remote location also prompted engineers to assemble the Correlator in four stages, with the first part of the machine arriving in Chile in 2008. "Reassembly and testing in Chile was about a 2-month effort for each of the four deliveries," says NRAO's Richard Lacasse, who helped lead the Correlator assembly, which finished its work late last month.
It was an ambitious undertaking. "In total there are a bit less than 3000 printed circuit cards of various complexities," Baudry notes. "And some of these cards may have a large number of individual electronic components." There are 32,768 custom chips performing multiplications, for example. Engineers have also had to stock up on spare parts in preparation for keeping the Correlator going over its estimated 20- to 30-year life span.
Could the Correlator become a model for other supercomputer efforts? Probably not, Baudry says, because the machine is so specific to ALMA's needs. But "perhaps the solutions and precautions devised to dissipate the energy released" by the computer's components, he says, "could be of interest for the future designers of general purpose supercomputers."