Sum-Product Networks (SPNs) and their credal counterparts are machine learning models that combine good representational power with tractable inference. Yet they often have thousands of nodes which result in high processing times. We propose the addition of caches to the SPN nodes and show how this memoisation technique reduces inference times in a range of experiments. Moreover, we introduce class-selective SPNs, an architecture that is suited for classification tasks and enables efficient robustness computation in Credal SPNs. We also illustrate how robustness estimates relate to reliability through the accuracy of the model, and how one can explore robustness in ensemble modelling.