AI BriefingGKE dynamic slicing for TPUs is now generally available, reducing startup times by up to 5x
·2026.11.09 09:00
Key point
The feature decouples logical TPU slice formation from physical nodepool provisioning using optical circuit switching to improve fleet utilization.
Details
- GKE dynamic slicing is now generally available, reducing AI workload startup times by up to 5x and workload recovery by 4.5x compared to static nodepools.
- The feature decouples logical TPU slice formation from physical nodepool provisioning, eliminating the need to manually delete and recreate nodepools for topology changes.
- Dynamic super-slicing uses TPU optical circuit switching (OCS) to stitch 16-VM TPU cubes into larger slices, while sub-slicing partitions a single cube into smaller, independent slices for multi-tenant workloads.
- Hardware failures are isolated to a single TPU cube, allowing the scheduler to swap unhealthy cubes in one to two minutes and reform slices using healthy partitions.
- Availability is limited to TPU All Capacity under reservation mode in all regions where TPUs are available, requiring a one-time pre-provisioning of 16-VM nodepools.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.