Google made Cloud Pathways generally available on Google Kubernetes Engine (GKE) on Nov. 9 in supported regions, giving cloud customers software to coordinate large AI workloads across its tensor processing units (TPUs).
The release covers GKE regions that support TPU v5e, TPU v5p, Trillium (TPU v6e) or Ironwood (TPU 7x), Google said. Google describes Cloud Pathways as a distributed machine-learning runtime designed to coordinate jobs across many accelerators.
A JAX process can coordinate devices spread across large TPU deployments, according to Google. JAX is Google's machine-learning framework. Google says Google DeepMind built Pathways to train and serve its leading AI models, including Gemini.
Google says traditional setups that rely on multiple controllers become harder to program and manage as machine-learning jobs expand across thousands of accelerators. The company presents Pathways as a way to address that coordination challenge for large workloads.
Google's GKE materials cover batch training, interactive workloads and inference distributed across multiple TPU hosts. For batch submissions, Google's guide demonstrates the Cluster Toolkit command gcluster job submit. The Google-maintained PathwaysJob repository says its API has been deprecated in favor of direct JobSets.
Google also documents controls for training interruptions. Suspend-resume supports planned pauses, though restoration depends on the workload. Elastic training addresses hardware failures and TPU slice preemptions; Google says jobs can be resized to use the healthy slices that remain.
