Google Makes Pathways on Google Kubernetes Engine Generally Available

Google made Pathways on Google Kubernetes Engine generally available Nov. 9, giving cloud customers access to software designed to coordinate large AI jobs across hundreds to thousands of TPU chips.

Google built Pathways to train large models, including Gemini. On Google Kubernetes Engine (GKE), a single JAX client can direct workloads across hundreds to thousands of Google's tensor processing unit (TPU) chips. JAX is a machine learning framework.

Pathways supports batch training and inference, interactive workloads and inference for large models running on multiple machines. GKE is Google's managed Kubernetes service, which provides the environment for deploying those workloads.

For long training runs, a suspend-resume feature lets a job pause and continue later. Pathways also supports elastic training for projects whose computing needs change during a run. Customers can submit and manage workloads through the PathwaysJob API.