far from the fastest - our work on Presto on GPUs would like a word ;-)