# NVIDIA Snapshot Cuts Kubernetes GPU Inference Starts to Seconds By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/54180 Published: 2026-11-11T15:57:47.000Z Updated: 2026-11-11T15:57:47.000Z Section: Technology > Snapshot preserves initialized CPU and GPU memory so new containers can skip setup steps. Tests on NVIDIA HGX B300 systems showed restore times of 3.6 to 42.7 seconds. NVIDIA introduced Snapshot, an open-source Kubernetes component, to restore initialized GPU inference workloads in seconds and reduce cold-start delays that can leave new capacity waiting minutes. Snapshot is part of NVIDIA’s DSX AI Factory Platform. It saves a workload’s CPU and GPU memory after initialization, then restores that state in a new container on compatible infrastructure. The container can skip repeated model-weight loading, kernel compilation and CUDA graph capture. Operators can save a running workload with a PodSnapshot or use a SnapshotJob to capture it immediately after initialization. A new pod can reference the saved snapshot. The component supports workloads that use multiple GPUs on one node, including setups using tensor, data and expert parallelism. On NVIDIA HGX B300 hardware, GLM-5.3 pod readiness was about 31 seconds, compared with roughly eight minutes without Snapshot. On the same platform, vLLM startup times ranged from six to 34 seconds, down from two to 8.5 minutes. Across tested configurations, restores took 3.6 to 42.7 seconds, an eightfold to 36-fold speedup over cold starts. The largest tested mixture-of-experts models used 1 TB checkpoints across eight GPUs and restored in less than 43 seconds, compared with cold starts of six to 19 minutes. These measurements cover tested configurations; they do not establish restore times for every model or infrastructure setup. Initially developed for NVIDIA Dynamo inference workers, Snapshot is available for use with other inference engines, schedulers and platforms. It supports vLLM, SGLang and NVIDIA TensorRT LLM. NVIDIA’s [Mission Control 2.5 Kubernetes recovery update]() addresses automated responses to failures in Kubernetes workloads managed by Run:ai. ## Links in this article - [Mission Control 2.5 Kubernetes recovery update](https://www.tokenpost.com/news/technology/41379)