NVIDIA introduces Snapshot to restore initialized GPU workloads

NVIDIA introduced Snapshot, an open-source system that saves initialized GPU workloads and restores them for inference scaling and recovery. The company reported restore times of 3.6 to 42.7 seconds across tested configurations, 8x to 36x faster than starting workloads from scratch. Snapshot captures CPU and GPU memory and restores it through Kubernetes, skipping repeated weight loading, compilation and graph capture. Examples support vLLM, SGLang and NVIDIA TensorRT LLM, including compatible workloads spanning multiple GPUs on one node. Restores require the same GPU model and driver. The measured restore path excludes infrastructure provisioning and container image loading.