Google Cloud GKE Dynamic Slicing Generally Available: Decouples TPU Nodepools, Cuts Startup Time by 5x
Google Cloud announced the general availability of GKE Dynamic Slicing, using optical circuit switching to enable on-demand TPU slice assembly and isolation, reducing AI workload startup times by up to 5x and accelerating recovery by 4.5x.
Google Cloud announced on November 10, 2026, that "Dynamic Slicing" is now generally available on its Kubernetes Engine (GKE). The feature addresses efficiency bottlenecks caused by the tight coupling between physical hardware configuration and logical compute topologies when running large-scale AI workloads on Tensor Processing Units (TPUs).
In traditional setups, managing TPU capacity required creating separate, static nodepools for each slice topology. This meant platform teams had to manually delete and recreate nodepools as workload needs changed, increasing complexity and creating large failure domains where a single hardware fault could crash an entire slice. GKE Dynamic Slicing decouples physical hardware provisioning from logical slice formation, leveraging TPU Optical Circuit Switching (OCS) technology for flexible topology management.
The feature operates in two modes: "super-slicing" and "sub-slicing." Super-slicing dynamically connects multiple TPU cubes (each consisting of 16 VMs) at scheduling time to form larger slices exceeding a single cube's size, catering to massive training demands. Sub-slicing partitions a single pre-provisioned TPU cube into multiple independent smaller slices, allowing concurrent, isolated execution of different workloads within the same physical cube. This eliminates the traditional "delete-and-create" nodepool cycle, enabling zero-downtime topology agility.
According to Google Cloud, compared to static nodepools, dynamic slicing reduces AI workload startup times by up to 5x, as the system instantly assembles logical slices from pre-provisioned nodepools instead of waiting for new deployments. Additionally, because failures are isolated to individual TPU cubes and unhealthy cubes can be swapped in one to two minutes, workload recovery speed improves by 4.5x. The feature also supports provisioning partially healthy nodepools, further enhancing overall fleet resource utilization.