Node Types

important

Due to high demand, A100 NVLINK (HGX) and H100 NVLINK (HGX) nodes are currently fully committed on client contracts, and are therefore not available for on-demand use. We recommend a conversation with the CoreWeave team to build a strategic plan catered to your needs to make use of available infrastructure and to plan for your future capacity requirements. Contact CoreWeave Sales to get started.

CoreWeave offers a "Standard Instance," which is useful for most workloads per GPU type offered on CoreWeave Cloud. These instances are a starting point, but can be configured entirely to suit your use case or compute needs.

You can view Standard Instance configurations on our website's pricing page.

For more information about à la carte pricing of compute components on CoreWeave Cloud, see Resource Based Pricing.

Component availability

The following table outlines the physical limitation of how many GPUs are available per instance when customizing your instances on CoreWeave Cloud.

Class	Generation	VRAM	Max per Instance	Label
H100 HGX	Hopper	80 GB	8	H100_NVLINK_80GB
H100 PCIe	Hopper	80 GB	8	H100_PCIE
A100 HGX	Ampere	80 GB	8	A100_NVLINK_80GB
A100 HGX	Ampere	40 GB	8	A100_NVLINK
A100 PCIe	Ampere	40 GB	8	A100_PCIE_40GB
A100 PCIe	Ampere	80 GB	8	A100_PCIE_80GB
A40	Ampere	48 GB	8	A40
RTX A6000	Ampere	48 GB	8	RTX_A6000
RTX A5000	Ampere	24 GB	8	RTX_A5000
RTX A4000	Ampere	16 GB	7	RTX_A4000
Tesla V100 NVLINK	Volta	16 GB	8	Tesla_V100_NVLINK
RTX 5000	Turing	16 GB	4	Quadro_RTX_5000
RTX 4000	Turing	8 GB	7	Quadro_RTX_4000

important

If a workload requests more peripheral compute resources (vCPU, RAM) than offered in a standard instance size, additional costs will incur.

CPU availability

CPU-only nodes are best suited for tasks such as control-plane services, databases, ingresses and CPU rendering.

CPU Model	Max RAM per vCPU	Max vCPU per Workload	Label
Intel Xeon v3	4 GB	70	intel-xeon-v3
Intel Xeon v4	4 GB	60	intel-xeon-v4
Intel Xeon Ice Lake	4 GB	94	intel-xeon-icelake
Intel Xeon Scalable	6 GB	94	intel-xeon-scalable
AMD Epyc Milan	4 GB	46	amd-epyc-milan
AMD Epyc Rome	4 GB	46	amd-epyc-rome

note

Workloads without GPU requests are always scheduled on CPU nodes.

Requesting compute in Kubernetes

A combination of resource requests and node affinity is used to select the type and amount of compute for your workload. CoreWeave Cloud relies only on these native Kubernetes methods for resource allocation, allowing maximum flexibility. The label used to select the GPU type is gpu.nvidia.com/class. CPU type is selected using the label node.coreweave.cloud/cpu.

note

These labels are mutually exclusive - a specific CPU type cannot be explicitly selected for GPU nodes.

Example specs

Single A100 80GB

spec:
  containers:
  - name: example
    resources:
      limits:
        cpu: 15
        memory: 97Gi
        nvidia.com/gpu: 1

  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
        - matchExpressions:
          - key: gpu.nvidia.com/class
            operator: In
            values:
              - A100_PCIE_80GB

8x A100 NVLINK

spec:
  containers:
  - name: example
    resources:
      requests:
        cpu: 90
        memory: 700Gi
      limits:
        nvidia.com/gpu: 8

  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
        - matchExpressions:
          - key: gpu.nvidia.com/class
            operator: In
            values:
              - A100_NVLINK

A100 With Fallback To A40

spec:
  containers:
  - name: example
    resources:
      limits:
        cpu: 12
        memory: 24Gi
        nvidia.com/gpu: 1

  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
        - matchExpressions:
          - key: gpu.nvidia.com/class
            operator: In
            values:
              - A100_PCIE_40GB
              - A40
      preferredDuringSchedulingIgnoredDuringExecution:
        - weight: 20
          preference:
            matchExpressions:
              - key: gpu.nvidia.com/class
                operator: In
                values:
                  - A100_PCIE_40GB

16 Core Xeon v3/v4 CPU

spec:
  containers:
  - name: example
    resources:
      limits:
        cpu: 16
        memory: 48Gi

  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
        - matchExpressions:
          - key: node.coreweave.cloud/cpu
            operator: In
            values:
              - intel-xeon-v3
              - intel-xeon-v4

Single Epyc CPU

spec:
  containers:
  - name: example
    resources:
      limits:
        cpu: 1
        memory: 4Gi

  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
        - matchExpressions:
          - key: node.coreweave.cloud/cpu
            operator: In
            values:
              - amd-epyc-rome
              - amd-epyc-milan

note

Kubernetes allows resources to be scheduled with requests and limits. When only limits are specified, the requests are set to the same amount as the limit.