Google Cloud makes GKE dynamic slicing generally available for TPUs

[Google Cloud] Google Cloud said GKE dynamic slicing is now generally available, decoupling logical TPU slice formation from physical nodepool provisioning. According to Google Cloud, dynamic slicing reduces AI workload startup times by up to 5x and workload recovery by 4.5x compared with creating a static nodepool for each TPU slice. The feature uses TPU optical circuit switching to stitch together TPU cubes, supporting super-slicing beyond a single cube and sub-slicing a cube into smaller slices for multiple workloads within one nodepool. Google Cloud said the approach eliminates the delete-and-create nodepool cycle, limits large slice failures to a 16-VM 4x4x4 slice topology, swaps unhealthy cubes in one to two minutes and allows use of partially healthy nodepools.