In this video from PASC18, Fernanda Foertter from NVIDIA presents: Practical Scaling Techniques for Deep Learning.
“The need for large scale training of neural networks is stemming from the advent of ever growing labeled datasets in data science combined with the successes of deep learning at achieving super-human performance at pattern recognition tasks and others. Fast and powerful GP-GPU have enabled such trainings thanks to an impressive level of parallelisation of computation. There remain however large problems which may take days to weeks to converge. To this end, additional level of parallelisation across computing units are used for additional speed up. We present an overview of the practical techniques which can be used for scaling throughput of model training.”
Thanks to Rich Brueckner from insideHPC Media Publications for recording the video.