ETS - Why is a GPU needed for NMT ?


Why is a GPU needed for NMT LP's? Are the LP's "learning" new translations all the time? Otherwise I can't understand why a fully trained LP would need a GPU.
SDL ETS with NMT LPs requires GPUs due to complexity of encoding and decoding when using pre-trained model.
During translation, the model is not being trained.
During training of NMT model, there is even more computational power needed compared to when only translating strings.

SDL ETS with NMT LPs could work with CPUs only, but the performance would be ~50-200 words per minute regardless how many top performance CPUs you use and therefore SDL don't support CPU only implementations.

SDL ETS 8.0 supports NMT Language Pairs ver. 8.0 on Linux/Unix Operating systems only. Support for NMT LP ver. 8.0 on Windows Operating system will be scheduled for Q1/Q2 of 2019.