Amazon EKS GPU Sharing: AWS Guide Shows Pod Over 8 GiB Cap Ends Completed, Skipping OOMKilled Alerts
Reported by Tech Times1 source ↓
An AWS and NVIDIA guide published Nov. 19 tested GPU memory sharing on EKS. A pod that requested 9 GiB against an 8 GiB cap ended as Completed, not OOMKilled, so alerts for that Kubernetes status would miss the error in this test.
Key points
- Two pods each ran on one NVIDIA A10G with an 8 GiB memory limit
- Both test pods allocated 6,144 MiB on the same EKS node
- The over-limit pod reported an out-of-memory error and exited with code 0
- The guide says applications may need log or GPU-memory metric monitoring
Details
The guide tested NVIDIA’s Run:ai GPU Fractioning Operator and KAI Scheduler on EKS 1.36, using a g5.xlarge node in us-east-2. With sharing disabled, one pod ran and a second remained Pending; with fractioning enabled, both ran on the single A10G GPU. The AWS post’s four authors include two people from AWS and two from NVIDIA.
The over-limit test set a pod’s GPU memory limit to 8,192 MiB and had it request 9,216 MiB. The guide says the probe caught the allocation error and exited successfully, producing a Completed status rather than a Kubernetes OOMKilled status; the result depends on how an application handles the error. The authors also say the test did not assess fault containment and that GPU memory limits do not reserve a proportional share of compute.
Written from coverage as of
Images
- Tech Times
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh