https://hgpu.org/?p=13888
Caffe con Troll: Shallow Ideas to Speed Up Deep Learning