Radio telescopes that employ arrays of many antennas are in operation, and ever-larger ones are being designed and proposed. Signals from the antennas are combined by cross-correlation. While the cost of most components of the telescope is proportional to the number of antennas, N, the cost and power consumption of cross-correlation are proportional to N2, and dominate at sufficiently large N. As radio telescopes get larger, there is a need to provide digital-signal-processing electronics that are smaller and less power-hungry than would be implied by the extrapolation of existing designs.
The goal of this work was to develop a custom integrated circuit (IC) that performs one of the most power-consuming processes — correlation — using an efficient architecture. The IC performs digital cross-correlations for arbitrarily many antennas in a power-efficient way. It uses an intrinsically low-power architecture in which the movement of data between devices is minimized. In a large system, each IC performs correlations for all pairs of antennas, but for a portion of the telescope’s bandwidth. In this design, the correlations are performed in an array of 4,096 complex multiply-accumulate (CMAC) units. This is sufficient to perform all correlations in parallel for 64 signals. When N is larger, the input data are buffered in an on-chip memory, and the CMACs are re-used as many times as needed to compute all correlations.
The design has been synthesized and simulated so as to obtain accurate estimates of the IC’s size and power consumption. As of this writing, physical design (layout) and fabrication of prototypes remain to be done. The IC design provides a power-efficient means of computing all cross-correlations among many signals. The power efficiency is more than two orders of magnitude better than that of existing large correlators, and about a factor of 20 better than planned correlators based on future-generation field-programmable gate arrays (FPGAs). The IC is flexible in that it can be used to construct correlators for almost any number of antennas, although its efficiency is best if N is a multiple of 64.