https://hgpu.org/?p=17580
Distributed Training Large-Scale Deep Architectures