https://hgpu.org/?p=21431
Brief Announcement: On the Limits of Parallelizing Convolutional Neural Networks on GPUs