The multi-filter rotating shadowband radiometer (MFRSR) uses a silicon photodiode sensor to measure shortwave global and diffuse horizontal irradiance from which direct normal irradiance is calculated. Silicon sensors are rugged, stable, and have a fast time response. However, silicon sensors are both thermally and spectrally sensitive. In addition they, as do all pyranometric sensors, have an imperfect cosine response, especially at high solar-incidence angles. In the MFRSR two of these sources of error are minimized: the cosine response of the MFRSR is measured, and the acquired data are corrected accordingly; an automatic heater maintans the MFRSR detector at a constant temperature near 40 degrees C. This paper demonstrates that there is substantial agreement between first-class thermopile instruments and the MFRSR silicon sensor with the elimination of these two sources of error. Furthermore, this paper describes corrections, based on sky conditions, to lower the remaining errors. The data base for deriving and testing these corrections was collected in Albany, New York, during 1993. After correction equations are applied, the root-mean-square differences for 5 min averages of the global horizontal, diffuse horizontal, and direct normal irradiance are 8.8, 9.1, and 16.3 W/m(2), respectively. The differences in time response or time keeping between silicon and thermopile instruments may explain much of the remaining root-mean-square differences.