NVIDIA open-sources Snapshot for faster GPU restarts
NVIDIA has open-sourced Snapshot, a Kubernetes tool that checkpoints fully initialized GPU inference workloads and restores them 8–36x faster than a cold start, shortening scale-up and failure recovery.
NVIDIA has released Snapshot, an open-source checkpoint and restore system for GPU workloads on Kubernetes. It ships as part of the NVIDIA DSX AI Factory Platform. Snapshot targets the cold-start problem: getting an LLM replica ready to serve can take minutes, which slows scaling when demand spikes and recovery when a replica fails.
Snapshot captures a running workload's CPU and GPU memory as a portable artifact. A new Pod can then resume from that state on any node with the same GPU model and driver, skipping weight loading, kernel compilation and CUDA graph capture. Operators use Kubernetes custom resources (PodSnapshot, SnapshotJob) and a nvidia.com/restore-from annotation. The workload itself needs only a small pause/resume handshake. The tool was built for NVIDIA Dynamo inference workers, and the repository includes examples for vLLM, SGLang and TensorRT LLM. It also supports multi-GPU, single-node MoE deployments that use tensor, data/expert or decode context parallelism.
NVIDIA ran its benchmarks on B200 GPUs with VAST storage. Restores took 3.6 to 42.7 seconds across models ranging from 600 million parameters to MoE models with checkpoints of roughly 1 TB. For the largest models, cold starts took 6 to 19 minutes. Neither measurement includes container image pulls.
Two design partners reported their own results on HGX B300. Crusoe, which contributes to the project upstream, cut GLM-5.3 Pod readiness from about 8 minutes to 31 seconds. Saturn Cloud reduced vLLM startup from 2–8.5 minutes to 6–34 seconds.
The post does not give full figures for checkpoint time or storage overhead in the excerpt available. NVIDIA describes the checkpoint as a one-time cost for each model and configuration.