Instance types

AWS instance types offer varying resources and can be selected by labels in your NodePool spec.template.spec.requirements. The values provided in this reference reflect resources available after instance overhead has been subtracted, with the following assumptions:

blockDeviceMappings are not configured
amiFamily is set to AL2023

How Karpenter discovers instance types

Karpenter calls the EC2 DescribeInstanceTypes and DescribeInstanceTypeOfferings APIs to discover the available instance types and their offerings (availability zone, capacity type) in your AWS region. This information is cached and periodically refreshed. Karpenter uses this data to:

Match NodePool requirements against available instance types
Estimate instance costs for optimized bin-packing
Select the best-fit instance type for each batch of pending pods

Karpenter filters out instance types that are incompatible with your NodePool requirements before considering them as candidates. Instance types in unsupported availability zones or without the requested capacity type (on-demand or spot) are automatically excluded.

Well-known labels

Karpenter annotates every node it launches with a rich set of labels describing the instance type’s properties. You can use these labels in nodeSelector, nodeAffinity, or NodePool requirements to influence scheduling.

Core labels

Label	Description	Example
`node.kubernetes.io/instance-type`	Full EC2 instance type name	`m5.xlarge`
`kubernetes.io/arch`	CPU architecture	`amd64`, `arm64`
`kubernetes.io/os`	Operating system	`linux`
`topology.kubernetes.io/zone`	Availability zone	`us-east-1a`
`karpenter.sh/capacity-type`	Capacity type	`on-demand`, `spot`

Karpenter AWS instance labels

Label	Description	Example values
`karpenter.k8s.aws/instance-category`	High-level instance category	`a`, `c`, `m`, `r`, `g`, `p`, `inf`, `trn`
`karpenter.k8s.aws/instance-family`	Instance family	`m5`, `c6g`, `p4d`
`karpenter.k8s.aws/instance-generation`	Generation number	`5`, `6`, `7`
`karpenter.k8s.aws/instance-size`	Instance size	`large`, `xlarge`, `2xlarge`, `metal`
`karpenter.k8s.aws/instance-cpu`	vCPU count	`2`, `4`, `8`, `96`
`karpenter.k8s.aws/instance-cpu-manufacturer`	CPU manufacturer	`aws`, `intel`, `amd`
`karpenter.k8s.aws/instance-cpu-sustained-clock-speed-mhz`	Sustained CPU clock speed in MHz	`3400`, `2500`
`karpenter.k8s.aws/instance-memory`	Total memory in MiB	`4096`, `32768`
`karpenter.k8s.aws/instance-ebs-bandwidth`	EBS bandwidth in Mbps	`4750`, `19000`
`karpenter.k8s.aws/instance-network-bandwidth`	Network bandwidth in Mbps	`1250`, `25000`
`karpenter.k8s.aws/instance-hypervisor`	Hypervisor type	`nitro`, `xen`, “ (bare metal)
`karpenter.k8s.aws/instance-encryption-in-transit-supported`	Supports encryption in transit	`true`, `false`
`karpenter.k8s.aws/instance-local-nvme`	Local NVMe storage in GB (if present)	`75`, `900`, `3800`
`karpenter.k8s.aws/instance-capability-flex`	Flex capacity support	`true`, `false`

GPU and accelerator labels

For GPU and accelerator instances (e.g., p, g, inf, trn families), Karpenter also surfaces:

Label	Description	Example
`karpenter.k8s.aws/instance-gpu-name`	GPU model name	`a100`, `v100`, `t4`
`karpenter.k8s.aws/instance-gpu-manufacturer`	GPU manufacturer	`nvidia`, `aws`
`karpenter.k8s.aws/instance-gpu-count`	Number of GPUs	`1`, `4`, `8`
`karpenter.k8s.aws/instance-gpu-memory`	GPU memory in MiB	`16384`, `40960`
`karpenter.k8s.aws/instance-accelerator-name`	Accelerator name (e.g. for Inferentia/Trainium)	`inferentia`, `trainium`
`karpenter.k8s.aws/instance-accelerator-manufacturer`	Accelerator manufacturer	`aws`
`karpenter.k8s.aws/instance-accelerator-count`	Number of accelerators	`1`, `16`

Architecture options

Karpenter supports both x86-64 (amd64) and ARM (arm64) architectures. The architecture is surfaced on the standard kubernetes.io/arch label.

apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: arm64-pool
spec:
  template:
    spec:
      requirements:
        - key: kubernetes.io/arch
          operator: In
          values: ["arm64"]
        - key: karpenter.k8s.aws/instance-category
          operator: In
          values: ["c", "m", "r"]

ARM-based Graviton instances (arm64) typically offer better price/performance for compute-intensive workloads. Look for instance families ending in g (e.g., c6g, m7g, r8g) for Graviton instances.

Filtering instance types in NodePools

Use spec.template.spec.requirements to constrain which instance types Karpenter can select. Requirements use standard Kubernetes label selectors with In, NotIn, Exists, DoesNotExist, Gt, and Lt operators.

Select by family

spec:
  template:
    spec:
      requirements:
        - key: karpenter.k8s.aws/instance-family
          operator: In
          values: ["m5", "m6i", "m7i"]

Select by category and minimum CPU

spec:
  template:
    spec:
      requirements:
        - key: karpenter.k8s.aws/instance-category
          operator: In
          values: ["c", "m"]
        - key: karpenter.k8s.aws/instance-cpu
          operator: Gt
          values: ["7"]   # more than 7 vCPUs (i.e., 8+)

Exclude specific sizes

spec:
  template:
    spec:
      requirements:
        - key: karpenter.k8s.aws/instance-size
          operator: NotIn
          values: ["nano", "micro", "small", "medium", "large"]

Require local NVMe storage

spec:
  template:
    spec:
      requirements:
        - key: karpenter.k8s.aws/instance-local-nvme
          operator: Exists

GPU workloads

To schedule GPU workloads, express the GPU resource requirement on your pod and add a matching NodePool:

# Pod spec requesting a GPU
resources:
  limits:
    nvidia.com/gpu: "1"

# NodePool for GPU instances
spec:
  template:
    spec:
      requirements:
        - key: karpenter.k8s.aws/instance-gpu-manufacturer
          operator: In
          values: ["nvidia"]
        - key: karpenter.k8s.aws/instance-category
          operator: In
          values: ["g", "p"]

Instance family reference

The following tables give a representative overview of the instance families Karpenter can discover. This list is not exhaustive — Karpenter queries the EC2 API dynamically and will include all types available in your region.

Compute-optimized (c family)

Family	Architecture	Manufacturer	Notes
`c1`	amd64	intel	Previous generation
`c3`	amd64	intel	Previous generation
`c4`	amd64	intel	Previous generation
`c5`	amd64	intel	Nitro hypervisor
`c5a`	amd64	amd	Nitro, encryption-in-transit
`c5ad`	amd64	amd	Nitro, local NVMe
`c5d`	amd64	intel	Nitro, local NVMe
`c5n`	amd64	intel	Enhanced networking
`c6a`	amd64	amd	6th gen AMD
`c6g`	arm64	aws	AWS Graviton2
`c6gd`	arm64	aws	Graviton2, local NVMe
`c6i`	amd64	intel	6th gen Intel
`c7g`	arm64	aws	AWS Graviton3
`c7i`	amd64	intel	7th gen Intel

General-purpose (m family)

Family	Architecture	Manufacturer	Notes
`m1`, `m2`, `m3`	amd64	intel	Previous generation
`m4`	amd64	intel	Previous generation
`m5`, `m5a`, `m5d`	amd64	intel/amd	Nitro
`m6a`	amd64	amd	6th gen AMD
`m6g`	arm64	aws	AWS Graviton2
`m6i`	amd64	intel	6th gen Intel
`m7g`	arm64	aws	AWS Graviton3
`m7i`	amd64	intel	7th gen Intel

Memory-optimized (r family)

Family	Architecture	Notes
`r3`, `r4`	amd64	Previous generation
`r5`, `r5a`, `r5d`	amd64	Nitro
`r6a`	amd64	6th gen AMD
`r6g`	arm64	AWS Graviton2
`r6i`	amd64	6th gen Intel
`r7g`	arm64	AWS Graviton3
`r7i`	amd64	7th gen Intel
`x1`, `x1e`, `x2`	amd64	High memory

GPU and accelerator instances

Family	GPU/Accelerator	Use case
`g4dn`, `g5`	NVIDIA T4 / A10G	ML inference, graphics
`p3`, `p4d`, `p5`	NVIDIA V100 / A100 / H100	ML training
`inf1`, `inf2`	AWS Inferentia	ML inference
`trn1`	AWS Trainium	ML training

GPU instance types require you to use an AMI that includes the appropriate GPU drivers (e.g., the NVIDIA driver for NVIDIA GPUs). Ensure your EC2NodeClass amiSelectorTerms selects an AMI with the correct drivers installed.

Available resources per node

The resources reported per node reflect allocatable capacity after Karpenter subtracts instance overhead. The key resources are:

Resource	Description
`cpu`	Allocatable CPU in millicores
`memory`	Allocatable memory in MiB
`ephemeral-storage`	Ephemeral storage (default `17Gi`)
`pods`	Maximum number of pods supported
`vpc.amazonaws.com/pod-eni`	ENI-based pod IPs (Nitro instances only)

For example, a c5.xlarge reports:

Resource	Quantity
cpu	3920m
ephemeral-storage	17Gi
memory	6584Mi
pods	58
vpc.amazonaws.com/pod-eni	18

Allocatable CPU and memory are lower than the raw instance specs because Karpenter accounts for Kubernetes system overhead (kubelet, OS processes) and the VM memory overhead percent configured via VM_MEMORY_OVERHEAD_PERCENT.

Get Started

Concepts

Guides

Reference

Help

How Karpenter discovers instance types

Well-known labels

Core labels

Karpenter AWS instance labels

GPU and accelerator labels

Architecture options

Filtering instance types in NodePools

Select by family

Select by category and minimum CPU

Exclude specific sizes

Require local NVMe storage

GPU workloads

Instance family reference

Compute-optimized (c family)

General-purpose (m family)

Memory-optimized (r family)

GPU and accelerator instances

Available resources per node

Build docs developers (and LLMs) love

Get Started

Concepts

Guides

Reference

Help

​How Karpenter discovers instance types

​Well-known labels

​Core labels

​Karpenter AWS instance labels

​GPU and accelerator labels

​Architecture options

​Filtering instance types in NodePools

​Select by family

​Select by category and minimum CPU

​Exclude specific sizes

​Require local NVMe storage

​GPU workloads

​Instance family reference

​Compute-optimized (c family)

​General-purpose (m family)

​Memory-optimized (r family)

​GPU and accelerator instances

​Available resources per node

Build docs developers (and LLMs) love

How Karpenter discovers instance types

Well-known labels

Core labels

Karpenter AWS instance labels

GPU and accelerator labels

Architecture options

Filtering instance types in NodePools

Select by family

Select by category and minimum CPU

Exclude specific sizes

Require local NVMe storage

GPU workloads

Instance family reference

Compute-optimized (c family)

General-purpose (m family)

Memory-optimized (r family)

GPU and accelerator instances

Available resources per node