NVIDIA CUDA 13.4 在 Kubernetes 中实现显存强制的 GPU 共享
原标题：Memory-Enforced GPU Sharing in Kubernetes with NVIDIA CUDA 13.4
AI 摘要
NVIDIA 发布 CUDA 13.4，新增对应用透明的显存边界强制能力，支持精确显存额度、可增长区间以及按容器选择计算共享模式（time-slicing 或 sm-sharing）。同时推出开源 Kubernetes operator gpu-fractioning，负责在 GPU 节点上运行 MPS、应用各 Pod 容器的显存限制与计算模式，并按分片而非整卡统计用量。KAI Scheduler v0.18.0 成为首个集成该能力的调度器，通过 NvFractions 模式写入 Pod 注解，实现 GPU 分片共享。
正文节选
Many AI workloads need only a fraction of a GPU. Giving each one an entire device can leave capacity idle; sharing GPUs lets platform teams serve more workloads with the infrastructure they already have. NVIDIA GPUs already support several approaches to sharing. Multi-instance GPU (MIG) splits a GPU into hardware-isolated instances that are transparent to the application, time-slicing lets several processes take turns without memory enforcement, and Multi-Process Service (MPS) lets them run conc