AWS eks: Update EKS ML AMI docs: add G7 support and NVIDIA driver selection
Summary
Adds the g7 instance family to the AL2023 x86_64 NVIDIA AMI support table, removes the prior warning that G7 requires a custom AMI with NVIDIA driver 595, and documents that a single AMI now ships two NVIDIA driver versions (580 LTS and 595 production) with automatic per-GPU driver selection at first boot. Also updates kernel version support (6.12 for K8s 1.33-1.35, 6.18 for 1.36+) and refreshes links to the renamed install-nvidia-drivers.sh and resolve-nvidia-driver.sh scripts.
Security assessment
The diff is purely about GPU instance type support, NVIDIA driver versioning, kernel versions, and script link updates for EKS ML-optimized AMIs. It contains no authentication, encryption, credential, IAM, or vulnerability-fix content, so it is not security-related.
Evidence
+AL2023 x86_64 NVIDIA | p6-b300, p6-b200, p5, p5e, p5en, p4d, p4de, p3, p3dn, g7, g7e, gr6, g6, g6e, g6f, gr6f, g5, g4dn
Diff
diff --git a/eks/latest/userguide/ml-eks-optimized-ami.md b/eks/latest/userguide/ml-eks-optimized-ami.md index 15f331523..6181426cc 100644 --- a//eks/latest/userguide/ml-eks-optimized-ami.md +++ b//eks/latest/userguide/ml-eks-optimized-ami.md @@ -21 +21 @@ EKS AMI variant | EC2 instance types -AL2023 x86_64 NVIDIA | p6-b300, p6-b200, p5, p5e, p5en, p4d, p4de, p3, p3dn, g7e, gr6, g6, g6e, g6f, gr6f, g5, g4dn +AL2023 x86_64 NVIDIA | p6-b300, p6-b200, p5, p5e, p5en, p4d, p4de, p3, p3dn, g7, g7e, gr6, g6, g6e, g6f, gr6f, g5, g4dn @@ -28,6 +27,0 @@ Bottlerocket x86_64 aws-k8s | inf1, inf2, trn1, trn2 -###### Important - -The G7 EC2 instance type requires NVIDIA driver version 595 or later. The EKS-optimized accelerated AMIs currently include NVIDIA driver version 580, which does not support G7 instances. To use G7 instances with Amazon EKS, you must build a custom AMI with NVIDIA driver version 595 using the [EKS AMI build scripts](./eks-ami-build-scripts.html). - -If you are using Karpenter for node provisioning and auto-scaling with EKS, it is recommended to exclude the `g7` instance family from your NodePools that use automatic AMI selection. See [`amiSelectorTerms`](https://karpenter.sh/docs/concepts/nodeclasses/#specamiselectorterms) in the Karpenter documentation for how to configure your Karpenter NodeClass to use your custom AMI for `g7` instances. - @@ -71 +65,5 @@ For details on the NVIDIA CUDA user mode driver and the CUDA runtime/libraries u -The EKS-optimized AL2023 NVIDIA AMIs support kernel 6.12 for Kubernetes versions 1.33 and above, and the NVIDIA driver 580 version for all Kubernetes versions. The NVIDIA 580 driver is required to use CUDA 13+. +The EKS-optimized AL2023 NVIDIA AMIs support kernel 6.12 for Kubernetes versions 1.33 through 1.35 and kernel 6.18 for Kubernetes version 1.36 and later. A single AMI includes two NVIDIA driver versions. + +Each node automatically selects the appropriate driver for its GPU at first boot and commits to that version for the lifetime of the node. The `g7` instance type requires NVIDIA driver version 595 or later, while all other supported GPUs (including `g7e`) require NVIDIA driver version 580. Each node selects the driver automatically based on its GPU. You don’t need to configure the driver version. + +The LTS (long-term support) driver version is currently 580, and the production branch driver version is currently 595. The NVIDIA 580 driver is required to use CUDA 13+. @@ -73 +71 @@ The EKS-optimized AL2023 NVIDIA AMIs support kernel 6.12 for Kubernetes versions -See the EKS-optimized [AL2023 releases](https://github.com/awslabs/amazon-eks-ami/releases) on GitHub for details of the component versions included in the AMIs. See the EKS AL2023 NVIDIA AMI [installation script](https://github.com/awslabs/amazon-eks-ami/blob/main/templates/al2023/provisioners/install-nvidia-driver.sh) and [kernel loading script](https://github.com/awslabs/amazon-eks-ami/blob/main/templates/al2023/runtime/gpu/nvidia-kmod-load.sh) for details on how the EKS AMIs configure the NVIDIA dependencies. You can find the list of installed packages and their versions on a running EC2 instance with the `dnf list installed` command. +See the EKS-optimized [AL2023 releases](https://github.com/awslabs/amazon-eks-ami/releases) on GitHub for details of the component versions included in the AMIs. For details on how the EKS AMIs configure the NVIDIA dependencies and select the driver version at boot, see the EKS AL2023 NVIDIA AMI [installation script](https://github.com/awslabs/amazon-eks-ami/blob/main/templates/al2023/provisioners/install-nvidia-drivers.sh) and [driver resolution script](https://github.com/awslabs/amazon-eks-ami/blob/main/templates/al2023/runtime/gpu/resolve-nvidia-driver.sh). You can find the list of installed packages and their versions on a running EC2 instance with the `dnf list installed` command. @@ -77 +75 @@ When building custom AMIs with the EKS-optimized AMIs as the base, it is not rec -When building custom AMIs for GPU instances, it is recommended to build separate custom AMIs for each instance type generation and family that you will run. The EKS-optimized accelerated AMIs selectively install drivers and packages at runtime based on the underlying instance type generation and family. For more information, see the EKS AMI scripts for [installation](https://github.com/awslabs/amazon-eks-ami/blob/main/templates/al2023/provisioners/install-nvidia-driver.sh) and [runtime](https://github.com/awslabs/amazon-eks-ami/blob/main/templates/al2023/runtime/gpu/nvidia-kmod-load.sh). +When building custom AMIs for GPU instances, it is recommended to build separate custom AMIs for each instance type generation and family that you will run. The EKS-optimized accelerated AMIs selectively install drivers and packages at runtime based on the underlying instance type generation and family. For more information, see the EKS AMI scripts for [installation](https://github.com/awslabs/amazon-eks-ami/blob/main/templates/al2023/provisioners/install-nvidia-drivers.sh) and [runtime](https://github.com/awslabs/amazon-eks-ami/blob/main/templates/al2023/runtime/gpu/resolve-nvidia-driver.sh). @@ -108 +106 @@ For details on the NVIDIA CUDA user mode driver and the CUDA runtime/libraries u -See the Bottlerocket Version Information in the [Bottlerocket documentation](https://bottlerocket.dev/en/) for details on the installed packages and their versions. The EKS-optimized Bottlerocket NVIDIA AMIs support kernel 6.12 for Kubernetes versions 1.33 and above, and the NVIDIA driver 580 version for Kubernetes versions 1.34 and above. The NVIDIA 580 driver is required to use CUDA 13+. +See the Bottlerocket Version Information in the [Bottlerocket documentation](https://bottlerocket.dev/en/) for details on the installed packages and their versions. The EKS-optimized Bottlerocket NVIDIA AMIs support kernel 6.12 for Kubernetes versions 1.33 through 1.35 and kernel 6.18 for Kubernetes version 1.36 and later. These AMIs include the NVIDIA driver 580 version for Kubernetes versions 1.34 and above. The NVIDIA 580 driver is required to use CUDA 13+.