メインコンテンツまでスキップ

GPU Software on Kubernetes for NVIDIA GPUs (H100, H200, L4, RTX PRO 6000)

Overview​

Managed Kubernetes with GPU on FPT Cloud supports NVIDIA H100, H200, L4, and RTX PRO 6000. This page covers the GPU Software option (GPU Operator), the GPU Driver Installation Types (Pre-install, User-install, and Managed-install), the supported driver versions — up to 615.71.09 (CUDA 13.4) — and how the driver type is shared between worker groups.

1. Supported GPUs and driver versions​

What you can use depends on the GPU card in your worker group — check this table before planning your cluster.

Driver versions by card and installation type​

Each cell shows the driver version and its CUDA version. For Managed-install, the table shows the newest version; older versions are also available — see GPU Driver Installation Type.

CardPre-installManaged-installUser-install
L4595.91.07 (CUDA 13.2)615.71.09 (CUDA 13.4)Supported
RTX PRO 6000580.82.07 (CUDA 13.0)615.71.09 (CUDA 13.4)Supported
A100535.54.03 (CUDA 12.2)Not supportedSupported
A30535.54.03 (CUDA 12.2)Not supportedSupported
H100570.133.20 (CUDA 12.8)615.71.09 (CUDA 13.4)Supported
H200570.133.20 (CUDA 12.8)615.71.09 (CUDA 13.4)Supported
注記

A100 and A30 don't support Managed-install, so the newer driver versions in this guide — including 615.71.09 — aren't offered for these cards.

2. GPU Software: GPU Operator​

GPU Operator installs and manages the NVIDIA software on your GPU nodes for you. To use H100/H200, pick a Kubernetes version that supports GPU Operator (v1.33 and v1.34) — other cards don't need this. When you create a cluster with one of these versions, the GPU Software section includes a GPU Operator option next to the existing options.

警告

Managed-install drivers require GPU Operator. If you select another GPU Software option, the Managed-install option doesn't appear.

警告

If you don't configure GPU Software, you must install GPU Operator yourself to run containers that use GPU resources, and install Network Operator to use GPUDirect RDMA over InfiniBand.

In the GPU Information section of the cluster settings, you set:

  • GPU Software: GPU Operator.
  • Software version: the GPU Operator version to install — the console lists the available versions, for example v26.7.2.
  • Cluster mig strategy: how MIG splits your GPUs. Only relevant for cards that support MIG.

3. GPU Driver Installation Type​

The installation type decides who installs the NVIDIA driver on your nodes and whether you can change its version later. Managed-install is the only type that lets you change the driver version after the cluster is created, and the only one that offers the newest versions.

TypeDescriptionUpdate version after cluster creation?Notes
Pre-installThe driver is pre-installed at the version you selectNoIf the version is deprecated, the corresponding worker loses support. The newest versions, such as 615.71.09, aren't available yet
User-installYou install the driver manuallyNoIncompatibilities may occur, and FCI doesn't support such cases
Managed-installThe driver is installed automaticallyYesRequires GPU Operator. Applies to H100, H200, L4, and RTX PRO 6000

Driver versions supported for Managed-install​

VersionCUDA
535.247.0112.2
550.163.0112.4
570.158.0112.8
580.82.0713.0
595.71.0513.2
615.71.0913.4

For RTX PRO 6000, the list starts at 580.82.07.

Notes on upgrading Managed-install drivers​

  • The system waits for all GPU workloads on the node to be undeployed before it upgrades the driver.
  • To start a manual upgrade, go to the fptcloud-gpu-operator namespace and delete the nvidia-driver-* pods on the corresponding node.
  • Changing the driver version of a running worker group interrupts GPU workloads on the node being upgraded — plan it in a maintenance window. The expected downtime is under 5 minutes.

4. Base and non-base worker groups​

The base worker group sets the driver type for the other worker groups in the cluster, so decide its driver type first.

警告

The base worker group doesn't support Managed-install. Use Pre-install or User-install for it.

How the driver type is shared between worker groups​

The driver type of the base worker group sets the driver type of the non-base worker groups:

Base worker group driverGPU of the non-base worker groupDriver type assigned
Pre-installAnyPre-install
User-installAnyUser-install

5. DRA on Kubernetes v1.34​

DRA (Dynamic Resource Allocation) is a Kubernetes feature, generally available from v1.34, that lets workloads request GPU resources more flexibly than the traditional device plugin method (nvidia.com/gpu). With DRA you can request GPUs by attribute such as driver version or memory, share one GPU across several containers, or allocate GPUs on demand. This improves GPU utilization and reduces cost.

To use DRA on Managed Kubernetes with GPU, you need:

  • A cluster that runs Kubernetes v1.34 or later.
  • The Managed-install driver type with a major version of 570 or higher — the driver versions 570.158.01, 580.82.07, 595.71.05, and 615.71.09 all meet this requirement.
  • kubectl installed on your machine.

For detailed installation and usage steps, see Using DRA for GPU.