The dynamical and radiative feedbacks from the deep convection over the tropical Pacific are quantified using ENSO signal in that region for both the observed system as well as for 16 climate models. Different from a previous analysis, we recognize the nonlinear relationship between deep convection and SST over that region, and perform the evaluation using the data from the warm phase and the cold phase separately. We also employ a much longer dataset than the previous analysis. While the results confirm the previous finding that most models underestimate the cloud albedo feedback and overestimate the water vapor feedback, we also show that the discrepancies mainly come from the warm phase, underscoring deep convection as a major source of error. The models have feedbacks of comparable magnitude and similar spatial pattern to the observation in the cold phase. Detailed examination of the cause of the weaker feedback from cloud albedo in the models reveals that the bias is linked to a weaker relationship between the short-wave cloud forcing and the precipitation in these models. In addition, a systematic feedback bias from the latent heat flux is revealed: the models tend to have a too strong positive feedback of latent heat flux over the central-western Pacific. The results suggest that the deficiency in the atmospheric feedbacks, particularly those from the deep convection, is a possible cause for the excessive cold-tongue in coupled models.