https://hgpu.org/?p=18166
mu-cuDNN: Accelerating Deep Learning Frameworks with Micro-Batching