The generalization of approaches based on deep learning and more recently the explosion of uses related to generative AI have established Nvidia GPUs as accelerators of first choice for these uses. The programming model offered by the CUDA API nevertheless involves concessions on the reproducibility of certain calculations. This talk will immerse you in the world of Nvidia GPUs by discussing what they are, how they are programmed and why some calculations do not always give the same result via a concrete example.