One client coordinates work across TPU slices
Google Cloud announced general availability of Pathways on Google Kubernetes Engine on November 9. Pathways is a runtime for teams using JAX to run machine-learning jobs across Cloud TPU resources. The release gives those teams another way to coordinate work when it grows beyond a single group of accelerators.
In the configuration Google documents, one JAX client submits work to a proxy that coordinates execution across TPU workers. That arrangement can let a team manage a job from one client even when its computation spans several slices. Google’s documentation distinguishes the CPU resources needed by the proxy from the accelerator resources used to execute the job.
Check hardware and scheduler before planning a rollout
Google lists TPU v5e, v5p, Trillium v6e and Ironwood TPU 7x for the generally available GKE release. The release announcement says the published container images support JAX 0.10.1 or later. It ties geographic availability to regions supporting those TPU configurations on GKE. The TPU location table can help teams identify candidate zones. Quota and capacity for a particular project still need verification.
Google describes the Slurm deployment on Cluster Director as a preview. The sources reviewed give no separate Pathways price for this launch. For a team comparing costs, CGH would account for the CPU and TPU resources in its own deployment plan. CGH did not provision resources or test job performance.