This paper describes a new approximate implementation of Spreading Activation (SA) for knowledge selection in very large datasets. SA is used to prime relevant knowledge domains and reduce considerably the graph queried and therefore the query time. The method is based on the representation of the dataset as a sparse matrix of integers and the application on the corresponding graph of fast path searching algorithm which counts the number of times a node is reached following independent paths. The algorithm is implemented and tested on a CUDA enabled GPU on a dataset containing about 100 million of nodes and 850 million of statements. The numerical evaluation indicates that the approximate SA mechanism proposed is quite promising for real time applications achieving the activation of about 64 million nodes and 374 million of statements in about 5.5 seconds.