Solving the corner-turning problem for large interferometers

Abstract

The so-called corner-turning problem is a major bottleneck for radio telescopes with large numbers of antennas. The problem is essentially that of rapidly transposing a matrix that is too large to store on one single device; in radio interferometry, it occurs because data from each antenna need to be routed to an array of processors each of which will handle a limited portion of the data (say, a frequency range) but requires input from each antenna. We present a low-cost solution allowing the correlator to transpose its data in real time, without contending for bandwidth, via a butterfly network requiring neither additional RAM memory nor expensive general-purpose switching hardware. We discuss possible implementations of this using FPGA, CMOS, analog logic and optical technology, and conclude that the corner-turner cost can be small even for upcoming massive radio arrays.