Ska and Evla Computing Costs for Wide Field Imaging

Abstract

I investigate the problem of high dynamic range continuum synthesis imaging in the presence of confusing sources, using scaling arguments and simulations. I derive a quantified cost equation for the computer hardware needed to support such observations for the EVLA and the SKA. This cost has two main components – from the data volume, scaling as D⁻⁶ (where D is the antenna diameter), and from the non-coplanar baselines effect, scaling as D⁻², for a total scaling of D⁻⁸. A factor of two in antenna diameter thus corresponds to 12 years of Moore’s law (18 month doubling time) cost reduction in computing hardware. For a SKA built with 12.5 m antennas observing with 1 arcsecond at 1.4 GHz, I find the computing load to be about 150 Petaflops (costing about $500 million in 2015). For 25 m antennas, the load is about 256 times lower, costing $2 million in 2015. This new cost equation differs from that of Perley and Clark (2003), which has scaling as D⁻⁶. This is because I find that the excellent Fourier plane coverage of the small antenna design does not significantly change the convergence rate of the Clean algorithm, which is already satisfactory in this regime.