NVIDIA Run:ai GPU Fractioning Now Supported on Amazon EKS
Open source NVIDIA Run:ai GPU Fractioning is now supported on Amazon EKS, allowing multiple memory-bounded Kubernetes Pods to share a single NVIDIA GPU. This update requires a custom driver setup but significantly increases GPU utilization for compatible workloads.
What changed?
NVIDIA Run:ai GPU Fractioning is now available as an open-source solution for memory-enforced GPU sharing on Amazon EKS. This capability relies on the GPU Fractioning Operator v0.1.7 and the KAI Scheduler, allowing multiple Kubernetes Pods to share a single NVIDIA GPU, with each Pod receiving an enforced GPU-memory contract. The validated environment uses an Amazon EC2 g5.xlarge instance (NVIDIA A10G GPU), Kubernetes 1.36, and a custom Amazon Linux 2023 AMI equipped with the NVIDIA R615 driver. This driver is not the default in EKS-optimized images and requires a custom build.

Why does it matter to an everyday developer?
GPU resources in Kubernetes clusters are often underutilized by workloads that do not fully consume the available GPU memory. Previously, a single Pod requesting GPU could prevent others from using unused GPU memory. With Run:ai GPU Fractioning, developers can now deploy multiple memory-bounded Pods on a single GPU, improving resource usage and enabling denser scheduling for interactive notebooks, model evaluation tasks, and small inference services. This is particularly relevant for teams sharing GPU nodes in Amazon EKS without Multi-Instance GPU (MIG) capability.
