Without proper control of numerical and methodological errors in theoretical predictions at the per mille level it is not possible to study the effect of input parameters in current hadron-collider measurements at the required precision. We present a new version of the parton-level code MCFM that achieves this requirement through its highly-parallelized nature, significant performance improvements and new features. An automatic differential cutoff extrapolation is introduced to assess the cutoff dependence of all results, thus ensuring their reliability and potentially improving fixed-cutoff results by an order of magnitude. The efficient differential study of PDF uncertainties and PDF set differences at NNLO, for multiple PDF sets simultaneously, is achieved by exploiting correlations. We use these improvements to study uncertainties and PDF sensitivity at NNLO, using 371 PDF set members. The work described here permits NNLO studies that were previously prohibitively expensive, and lays the groundwork necessary for a future implementation of NNLO calculations with a jet at Born level in MCFM.