We report the results of a large-scale, multi-year experimental evaluation of the System of Rice Intensification (SRI), an innovation that first emerged in Madagascar in the 1980s and has now diffused to more than 50 countries. Using a randomized training saturation design, we find that greater cross-sectional or intertemporal intensity of direct or indirect training exposure to SRI has a sizable, positive effect on Bangladeshi farmers’ propensity to adopt (and not to disadopt) SRI. We find large, positive and significant impacts of SRI training on rice yields and profits, as well as multiple household well-being indicators, for both trained and untrained farmers in training villages. Despite the significant farm-level impacts on rice productivity and labor costs, we find no evidence of significant general equilibrium effects on rice prices or wage rates. We also find high rates of disadoption, and clear indications of non-random selection into technology adoption conditional on randomized exposure to training, such that adopters and non-adopters within the same treatment arm experience similar outcomes. Rice yields, profits and household well-being outcomes do not, however, vary at the intensive margin with intensity of training exposure, a finding consistent with multi-object learning models.