Translate
Language Weaver Edge is deployed in Kubernetes as pods, running on standard CPU nodes. Translation Engine pods & Training Engine pods could optionally be deployed in Kubernetes on nodes with Nvidia GPUs. Helm charts provided by RWS will try to deploy Training Engines & Translation Engines on nodes with Nvidia GPUs if detected & specified in the “values.yaml”. No pods are scheduled on Kubernetes GPU nodes by default, unless specified in the “values.yaml”.
For Translation Engines, the cost of deploying on GPU nodes, tends to be comparable with deploying on non-GPU nodes. The additional complexity may not be warranted.
For Training Engines, GPU deployment offers significant speed gains versus CPU deployment (up to 10x faster).
Training Adaptable Language Pairs on GPUs is cost effective when a Language Pair takes unacceptably long time to complete a training on CPUs.
Advantages
- Training Engines could be considerably faster (up to 10x speedup).
Disadvantages
- Kubernetes cluster must have GPU nodes.
- Auto scaling Translation Engines is not yet supported on GPUs.
- The hardware cost of performing translations is usually comparable with CPU deployment.
Supported GPU Platforms
- The current release of Language Weaver Edge supports only Nvidia GPUs.
Configuration
- Adding below under Training Engine or Translation Engine Stateful Set configuration in “values.yaml” will force the helm deployment to use GPU nodes to deploy Training Engine or Translation Engine pods.
gpu: true