https://hgpu.org/?p=17288
Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour