Really appreciate the work put in by you guys to compare all these libraries across so many different models and configurations.
I had a minor nitpick. For Torch7, it seems you are using https://github.com/Element-Research/rnn.
However, this library, as far as I know, does not leverage CuDNN bindings or specialised CUDA kernels for RNNs, something which is par for the course for other libraries such as TF, CNTK, etc.
For Torch7, you should be using either:
These libraries have specialised CUDA kernels for RNNs, and should bring the RNN benchmarks for Torch7 up to a comparable position compared to other libraries being benchmarked.
Hi,
Really appreciate the work put in by you guys to compare all these libraries across so many different models and configurations. I had a minor nitpick. For Torch7, it seems you are using https://github.com/Element-Research/rnn. However, this library, as far as I know, does not leverage CuDNN bindings or specialised CUDA kernels for RNNs, something which is par for the course for other libraries such as TF, CNTK, etc. For Torch7, you should be using either:
These libraries have specialised CUDA kernels for RNNs, and should bring the RNN benchmarks for Torch7 up to a comparable position compared to other libraries being benchmarked.