How does TensorFlow handle multiple samples?

Question

Say the mini-batch has $N$ samples $(x, y)$,

how will tensorflow utilize this $N$ samples to train the network.

1. Will it do $N$ forward loop for each sample independently?
2. Will it do $N$ backward propagation and $N$ weights update for each sample independently?
3. Or will it average the loss from $N$ samples and then do only 1 weights update?

Answer 1

3. Loss is computed for each sample and then averaged over the entire mini-batch and the weights are updated once. Check this video (start from 6:11) for details.