CUDA 13.4 memory-enforced GPU sharing

Overview

An NVIDIA post dated October 27, 2026 describes memory-enforced GPU sharing in CUDA 13.4: a memory boundary that needs no application changes, giving each workload an exact memory amount (allocations above it fail), a guaranteed minimum with an optional burstable maximum, a choice of time-slicing or sm-sharing compute mode, and a view of only its own fraction. On Kubernetes the open-source GPU Fractioning Operator, in the KAI Scheduler GitHub organization, runs next to the NVIDIA GPU Operator, supervises MPS on each node and applies per-container limits. KAI Scheduler v0.18.0 or later integrates through its NvFractions mode via Pod annotations. NVIDIA provides no benchmarks.

This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.

Key features

• Hard per-workload GPU memory limits without code changes
• Guaranteed minimum with optional burstable maximum
• Time-slicing (default) or sm-sharing compute modes
• Open-source GPU Fractioning Operator working with the NVIDIA GPU Operator
• Requires KAI Scheduler v0.18.0+ with NvFractions enabled

Use cases

• Packing multiple small inference workloads onto one GPU safely
• Multi-tenant Kubernetes GPU clusters needing memory isolation

Pricing

Operator and KAI Scheduler are open source; CUDA 13.4 general availability date not stated.

Pricing reflects the source record and may change. Confirm current terms with the provider.

Dates and sources

Released

27 Oct 2026

Record updated

29 Oct 2026

Record last checked

29 Oct 2026

• developer.nvidia.com/blog/memory-enforced-gpu-sharing-in-kubernetes-with-nvidia-cuda-13-4/