Slice
Stay organized with collections
Save and categorize content based on your preferences.
Slice is a custom resource definition that lets you define and manage
accelerator slices in Google Kubernetes Engine (GKE). This resource lets you
group TPU partitions to form larger slices for distributed training workloads.
When you create or update a Slice custom resource, a validating webhook is
triggered to ensure that the resource specification is valid. This validation helps to ensure that only valid Slice definitions are
accepted by the Kubernetes API server.
The name of the Slice resource. It must adhere to the following rules to
ensure compatibility with underlying Compute Engine resource naming conventions:
Length: the name must be 49 characters or fewer. The
controller appends a hyphen and an 8-character cluster hash to create
Compute Engine resource names, which have a 63-character limit.
Format: it must match the following regular expression: ^[a-z]([-a-z0-9]*[a-z0-9])?$.
This match means the name has the following characteristics:
Must start with a lowercase letter.
Can contain only lowercase letters, numbers, and hyphens (-).
Must end with a lowercase letter or a number (it can't end with a hyphen).
metadata.annotations
optional
object
Annotations for the Slice resource.
slice.gke.io/retry-on-failure: "true"
If set to "true", the slice controller will retry the creation of the slice if it fails.
type
required
string
The type of accelerator for this slice. The type field must be one of the following supported values:
tpu7x
Any other value will be rejected.
topology
required
string
The topology for the accelerator slice. Requirements for the tpu7x accelerator topology include the following rules:
The topology must be a three-dimensional string in following format: AxBxC. For example, 4x8x8.
For dynamic sub-slicing (topologies smaller than 4x4x4), you can specify 2x2x1, 2x2x2, 2x2x4, or 2x4x4. Requires GKE version 1.36.0-gke.3712000 or later.
For dynamic super-slicing (topologies greater than or equal to 4x4x4), each dimension of the requested topology must be a multiple of four. Requires GKE version 1.35.2-gke.1842000 or later.
The dimensions must be sorted in non-decreasing order: A <= B <= C. For example, 4x8x4 is invalid; it should be 4x4x8.
The product of the dimensions (A*B*C) must not exceed 9,216.
partitionIds
required
string[]
A list of strings that identify the partitions making up the slice. All values within the partitionIds list must be unique. Duplicate partition IDs are not allowed. Requirements include the following:
For dynamic sub-slicing (smaller than 4x4x4), you must specify exactly one partition ID. Requires GKE version 1.36.0-gke.3712000 or later.
For sub-blocks equal to or larger than 4x4x4, the number of partitions is calculated based on the total number of chips, where each partition consists of 64 chips. The number of items in your spec.partitionIds list must exactly match the calculated number of partitions ((A*B*C) / 64).
Condition status. Possible values are True or
False.
conditions.reason
string
A short description of the lifecycle state of the slice.
Possible values include the following:
SliceNotCreated: the slice hasn't been created yet. The slice
controller is currently initializing and performing preflight checks
for slice formation.
SliceCreationFailed: creation failed because prerequisites
were not met (for example, required Compute Engine resources are missing)
or preflight checks failed. To automatically retry slice formation in this
state, configure the slice.gke.io/retry-on-failure: "true" annotation.
ACTIVATING: the slice is in the process of being formed
(stitched).
ACTIVE: the slice is fully formed, healthy, and ready to
execute workloads.
ACTIVE_DEGRADED: the slice is formed but includes degraded
cubes, supported by ICI resilience. Workloads can run, but performance
might be impacted. Dynamic super-slicing is resilient to single optical
circuit switch (OCS) failures. If a single OCS unit fails, all optical
links traversing that switch become unavailable, which causes all cubes
within the superpod to operate in a degraded state. This state applies only
to the top-level 4x4x4 topology. Smaller topologies don't
have a degraded state.
DEACTIVATING: the slice is in the process of being dismantled.
FAILED: the slice is no longer ready to execute workloads.
This state occurs if initial formation failed or if an active slice
experienced a critical software or hardware failure.
INCOMPLETE: for dynamic sub-slicing, this is the terminal step
before the resource is fully deleted. For dynamic super-slicing,
there are not enough available cubes to initiate super-slice formation.
[[["Easy to understand","easyToUnderstand","thumb-up"],["Solved my problem","solvedMyProblem","thumb-up"],["Other","otherUp","thumb-up"]],[["Hard to understand","hardToUnderstand","thumb-down"],["Incorrect information or sample code","incorrectInformationOrSampleCode","thumb-down"],["Missing the information/samples I need","missingTheInformationSamplesINeed","thumb-down"],["Other","otherDown","thumb-down"]],["Last updated 2026-07-31 UTC."],[],[]]