Sound sources localization (SSL) is a subject of active research in the field of multi-channel signal processing since many years, and could benefit from the emergence of data-driven approaches. In the present paper, we present our recent developments on the use of a deep neural network, fed with raw multichannel audio in order to achieve sound source localization in reverberating and noisy environments. This paradigm allows to avoid the simplifying assumptions that most traditional localization methods incorporate using source models and propagating models. However, for an efficient training process, supervised machine learning algorithms rely on large-sized and precisely labelled datasets. There is therefore a critical need to generate a large number of audio data recorded by microphone arrays in various environments. When the dataset is built either with numerical simulations or with experimental 3D soundfield synthesis, the physical validity is also critical. We therefore present an efficient tensor GPU-based computation of synthetic room impulse responses using fractional delays for image source models, and analyze the localization performances of the proposed neural network fed with this dataset, which allows a significant improvement in terms of SSL accuracy over the traditional MUSIC and SRP-PHAT methods.