Google Cloud announced general availability of GKE dynamic slicing for TPUs on November 9, 2026. It lets a workload get the TPU topology it requests by combining pre-provisioned pools or dividing one pool into smaller slices, rather than requiring a separate fixed pool for every topology. Google’s announcement and GKE’s technical documentation describe two modes: super-slicing and sub-slicing.

The key planning distinction is whether a workload needs TPU capacity combined across pools or partitioned within one pool.

Super-slicing combines pools; sub-slicing divides one

Super-slicing joins multiple pre-provisioned TPU pools to form a larger slice for a workload. Google’s documentation describes using the optical circuit switch (OCS) network fabric to extend the inter-chip interconnect (ICI) between pools.

Sub-slicing partitions a single pre-provisioned pool into smaller, independent slices. For Ironwood (TPU7x), GKE lists sub-slice topologies of 2×2×1, 2×2×2, 2×2×4 and 2×4×4. The documentation describes the pool’s base unit as a 4×4×4 sub-block. These are topology dimensions, not a measure of workload speed or capacity in isolation.

In both modes, GKE forms the logical slice topology when it schedules the workload. Its slice controller manages the Kubernetes Slice custom resource and the associated topology configuration. Google’s GKE concepts documentation explains the two modes and their topology requirements.

Check the documented hardware and GKE versions

Google’s detailed requirements specify Ironwood (TPU7x), a Standard GKE cluster on the Rapid release channel, Container-Optimized OS node images and TPU All Capacity reservations. The minimum GKE version differs by mode:

Slicing modeMinimum GKE version on Rapid
Super-slicing1.35.2-gke.1842000 or later
Sub-slicing1.36.0-gke.3712000 or later

The version distinction matters when planning a cluster: meeting the minimum for one mode does not meet the stated minimum for the other. Google lists these requirements in its GKE dynamic slicing documentation.

Set up the controller and scheduling components

Google’s setup guide covers enabling the GKE slice controller, provisioning TPU node pools and configuring workload scheduling. One documented path uses Kueue with Topology Aware Scheduling (TAS) to create Slice resources; a custom scheduler can manage those resources as well.

The setup guide gives different minimum component versions for each mode:

  • Super-slicing: Kueue v0.16.6 or later and JobSet v0.11.1 or later.
  • Sub-slicing: Kueue v0.18.2 or later and JobSet v0.12.0 or later. LeaderWorkerSet is also required for this mode.

JobSet is required for multislice workloads. For a platform team, the practical check is to match the intended slicing mode against the GKE version and scheduler-component requirements before configuring workloads.

Google’s reported performance figures

Google says dynamic slicing can make job startup up to 5× faster and recovery up to 4.5× faster compared with static node-pool approaches. For super-sliced workloads, Google also says an unhealthy TPU cube can be replaced in one to two minutes. These are Google’s reported comparisons and recovery figure. The announcement describes the results in the context of forming and recovering slices from pre-provisioned pools.