https://hgpu.org/?p=11941
One weird trick for parallelizing convolutional neural networks