NVIDIA introduced Snapshot, an open-source checkpoint and restore component for GPU workloads on Kubernetes, in a technical blog post published on 10 November 2026. The company reports restoring initialized inference replicas in 3.6–42.7 seconds across its tested configurations, an 8×–36× speedup over cold starts.[1]
Snapshot captures CPU and GPU memory after workload initialization. NVIDIA says a new Pod can resume from that artifact on a node with the same GPU model and driver, skipping weight loading, compilation and CUDA graph capture. Compatible workloads can span multiple GPUs within one node. The workload must also implement a small pause-and-resume handshake and meet checkpoint, restore and runtime compatibility requirements.[1]
The reported benchmarks used NVIDIA B200 GPUs, driver 615.71.09 and VAST NFS storage, with vLLM 0.31.0, SGLang 0.5.21 and TensorRT LLM 1.3.0. Cold-start timing ran from process launch until engine initialization returned, before inference; restore timing ran from the restore request until serving resumed. Infrastructure provisioning and container image loading are outside the measured restore path. NVIDIA also says the key-value cache was released before every checkpoint and excluded from the artifact.[1]
[1]: NVIDIA, “Start and Restore GPU Workloads in Seconds with NVIDIA Snapshot,” 10 November 2026. https://developer.nvidia.com/blog/start-and-restore-gpu-workloads-in-seconds-with-nvidia-snapshot
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
9d1f6998089855d842ff9d77429246dc1e1512a1b4dd5e9b33cd48ec32592e07