NVIDIA open-sources Snapshot to cut GPU workload startup to seconds

NVIDIA published Snapshot, an open-source checkpoint and restore system for GPU workloads on Kubernetes, as a component of its DSX AI Factory Platform. It saves a fully initialized workload, including CPU and GPU memory, as a portable artifact and restarts it on any node with the same GPU model and driver, so the workload resumes without reloading weights, recompiling kernels or recapturing CUDA graphs. Checkpoint and restore are driven by Kubernetes custom resources, and the repository ships examples for vLLM, SGLang and TensorRT LLM. On B200 GPUs with VAST NFS storage, NVIDIA measured restores from 3.6 to 42.7 seconds, including multi-GPU mixture-of-experts models with checkpoints up to 1 TB across eight GPUs. The code is available now.