Documentation Center

Technical requirements for language pair adaptation

Technical requirements in terms of PUs, GPU, CPU and OS to perform language pair adaptation.

Minimum system requirements for language pair adaptation

See Recommended system requirements for more information about the system requirements to perform language pair adaptation in Language Weaver Edge.

Processing units for language pair adaptation

Similarly to generic language pairs, 1 processing unit (PU) is required to run an adapted or auto adaptive language pair with a throughput of 2000 words per minute.

To run multiple adapted language pairs in parallel, you need as many PUs as language pairs. For example, if you create 2 adapted language pairs for English → French and would like to run them in parallel, you will need 2 PUs.

Language pair adaptation on GPU

The fastest way to perform adaptation is to use a GPU. Using a GPU allows completing the adaptation process in a few hours.

1 GPU can perform the adaptation of 1 language pair at a time.

During the adaptation, 1 GPU is used exclusively for the training of the language pair. It's not recommended to have any started translation engines using the GPU on the host while the training is taking place.

Once the adapted language pair has been created, it can be run on either CPU or GPU.

See Supported GPUs for more information about the GPUs that can be used in Language Weaver Edge.

Language pair adaptation on CPU

If you don't have a GPU, you can still adapt language pairs by enabling a training engine on a worker with CPU only.

The adaptation done on CPU provides identical output as if the training had been performed on GPU, but the process is slower. While a training on GPU can take between 1 and 5 hours, a training on CPU will take between 1 and 5 days, depending on the size of the training data.

OS for language pair adaptation

Language pair adaptation can be done in both Linux and Windows, using either CPU or GPU.

See Compatible operating systems for more information about the compatible operating systems.