https://hgpu.org/?p=11142
GPU Asynchronous Stochastic Gradient Descent to Speed Up Neural Network Training