Sum-product networks (SPNs) are a class of probabilistic models that were introduced for the first time in the early 2010s. SPNs belong to the intersection of deep models and probabilistic graphical models with the ability to guarantee tractable inference. SPNs can encode any probabilistic distribution in a directed acyclic graph, in which each non-leaf node is either a sum or a product node. This paper, introduces a new inference method in SPNs that can prune ineffective subnetworks for a query during inference. By keeping some additional information in sum nodes, we can prune a few nodes that have no effect on the output of the network for an input sample. Experimental results demonstrated that even 99% of the nodes could be pruned from some datasets without affecting output accuracy.