https://hgpu.org/?p=6760
Improving the speed of neural networks on CPUs