The open-source GPU Fractioning Operator pairs with KAI Scheduler to apply CUDA 13.4 memory boundaries per container.

NVIDIA introduced the open-source GPU Fractioning Operator under the KAI Scheduler organization on GitHub. The operator runs alongside the NVIDIA GPU Operator, configures Multi-Process Service on each GPU node, and applies per-workload memory limits and compute modes. CUDA 13.4 adds memory boundary enforcement that applications do not need to manage.

Platform teams already use MIG, time-slicing, and MPS to share NVIDIA GPUs. MIG splits a device into hardware-isolated instances, time-slicing lets processes take turns without memory enforcement, and MPS runs processes concurrently in separate address spaces. KAI Scheduler v0.18.0 adds an NvFractions mode that writes per-container Pod annotations for the operator to enforce.

Builders can request exact memory sizes, guaranteed minimums with burstable maximums, and sm-sharing or time-sharing compute modes through annotations. The operator measures memory and compute use per GPU fraction, which gives platform teams clearer capacity accounting than whole-device metrics. Administrators can enable NvFractions during KAI Scheduler installation with a Helm flag.

The GPU Fractioning Operator repository includes installation and configuration guidance, and the project maintains admin and user guides for the three GPU sharing mechanisms. Operators should watch how NvFractions annotations interact with MPS setup across mixed GPU fleets and how CUDA 13.4 memory boundaries affect burst behavior under production load.

What matters

• NVIDIA launched an open-source Kubernetes operator that sets up MPS and enforces GPU memory limits.
• Platform teams can pack more workloads onto existing GPUs without handing each job a full device.
• KAI Scheduler v0.18.0 adds NvFractions annotations, and admins enable the mode with a Helm flag.

Why it matters

KAI Scheduler v0.18.0 adds NvFractions annotations, and admins enable the mode with a Helm flag.

This GenAI News article was prepared in original wording using reporting and materials published by NVIDIA Developer Blog. Source reference: https://developer.nvidia.com/blog/memory-enforced-gpu-sharing-in-kubernetes-with-nvidia-cuda-13-4/.

Drafted by the GenAI News review pipeline.