One big caveat is that they use a parallel algorithm on a 44 cores machine to compete with a single GPU.
I can train a neural network on the GPU in my laptop but it has only 8 cores.
I can train a neural network on the GPU in my laptop but it has only 8 cores.
Also, I'm pretty sure the goal of this research was not that you can train a DNN on your laptop, but that there is more democratization in the distribution of training power. Maybe some day a mid sized organization will be able to train a competitive model on a buch of off the shelf (or rented) servers...