| Description: | Moore's law indicates technology will advance to provide the required computing power for SKA some day. The improvement of key algorithms does not grow with Moore's law. This is the consequence of their computational complexity (FFT~O(nlogn); Correlation~O(N^2) ... ), solving is minimally a cubic problem. The complexity depends also on the system archicture, at some point FX correlators outperform XF. The required processing power for calibration is not precisely known, we provide a guestimate here, but it is likely to be for more than the other processing steps according to it's scaling behavior. Current operation relies on monolithic packages (Miriad, Newstar, AIPS). A transition to distributed solvers is required. The calibration algorithms are far more difficult to parallelize and distribute. Moreover a dedicated machine based on FPGA or ASIC is not suitable for high-precision floating point computing. This clarifies the computing challenges we face. I suggest we estimate these guestimates to assess the importance of this challenge. |
