Accurate cloud top height retrievals from hyperspectral infrared (IR) sounder radiances are needed for weather and climate prediction. To account for the nonlinearity of the cloud parameters with respect to the IR radiances, a one-dimensional variational retrieval algorithm is used to derive the cloud top heights (CTHs) from the Atmospheric Infrared Sounder (AIRS) radiances on a single field-of-view basis. The CTHs are evaluated by comparison with the measurements from radar and lidar instruments onboard the Earth Observing System (EOS) CloudSat and Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) satellites. Using the retrievals from a global 3 day dataset, it is found that the variational algorithm compared with the regression algorithm could reduce the variability of the difference between the AIRS and active measurements by 1 km. And the biases of AIRS CTHs range from +1.5 to -1.4 km and from +1.6 to -3.8 km, depending on the Cloud Profiling Radar (CPR) and CALIPSO CTHs between 3 and 18 km, respectively. Globally, the AIRS CTH is overestimated (underestimated) when the CTH from active measurements is below (above) 5 km. The bias decreases from -1.9 to -0.8 km, and the variability decreases from 2.8 to about 1.6 km with the increase of the CALIPSO cloud optical thickness from 0.1 to 2.5. It also reveals that the AIRS CTHs agree better with the CPR than the CALIPSO.