PyTorch is basically a direct Torch reimplementation in Python with some architectural improvements.
I am not aware of any benchmarks but I would give a slight edge to PyTorch since it receives max devs attention today.
I don't think that any speed difference if it exists will make a huge difference provided you have a good GPU which you should anyway.
What would make a difference is the ability to move you neural network into production and it is much more difficult to do with Torch/Lua unfortunately.
PyTorch is not production friendly either but at least there are converters in development.
I really hoped that Torch would invite academia into Lua world and the whole Lua ecosystem receives a huge benefit from new users coming (I learnt Lua because of Torch).
But luajit and 5.3 disconnection put an end to these hopes and persuaded the whole dev team to migrate to Python asap.
I will continue using Torch because of Lua but afraid that eventually I will join the flock soon.
