I don't know much about this (it's not my area of expertise), but I thought this G+ post was interesting: https://plus.google.com/u/0/+VincentVanhoucke/posts/6R

I don't know much about this (it's not my area of expertise), but I thought this G+ post was interesting: https://plus.google.com/u/0/+VincentVanhoucke/posts/6RQmgqcmx2d https://plus.google.com/u/0/+VincentVanhoucke/posts/6RQmgqcm... It says that a lot of the reason TensorFlow initially lagged in performance is because a lot of those performance issues only manifested under NVCC, whereas they had been using GPUCC internally.