Crusoe says it cut the time to get a GLM-5.3 inference pod ready on NVIDIA HGX B300 from about eight minutes to 31 seconds using NVIDIA Snapshot. The key finding is that local NVMe storage, rather than a cleverer checkpoint tool, did most of the heavy lifting.
Crusoe Watch analysis
What happened
In a benchmark published on 10 November, Crusoe says Snapshot restored 594GB of GPU state from local NVMe in 15 seconds, at about 40GB/s. Restoring the same state from NFS took 33.3 seconds, roughly 18GB/s. The benchmark used Snapshot release v0.2.0-rc.2, with Crusoe’s code in a project pull request.
Crusoe reports that total agent restore time fell from 41.4 seconds with NFS to 22.9 seconds with NVMe. The full journey from pod creation to readiness dropped from 50 seconds to 31. Crusoe also says its restore-failure fixes have been merged into the open-source Snapshot project.
Why it matters
For inference operators, a GPU sitting idle while its workload reloads is expensive furniture. Faster recovery and scale-up could help available capacity start serving requests sooner, especially when demand spikes or a worker fails.
The result also points to a practical limit: the benchmark shows a substantial gain on this hardware and setup, not a universal 15-fold speed-up for every model or deployment. Storage choice is central to the result, and the remaining delay includes process restoration and pod startup.
Our read
This is a useful operational benchmark, not just another promise that AI infrastructure will become more efficient. Crusoe’s numbers make the bottleneck visible: once a large GPU checkpoint is involved, where it lives can matter more than the restore machinery around it. Operators should treat the result as a reason to test their own storage path, not as a performance guarantee.
What to watch
- Whether independent operators reproduce the NVMe-versus-NFS results on other workloads.
- How much the remaining restore time can be reduced beyond storage improvements.
- Whether the merged Snapshot fixes and performance work appear in later stable releases.
Discussion spark: For AI inference infrastructure, should operators prioritise faster recovery from local NVMe, even if that adds storage cost and complexity, or favour simpler shared storage?
Sources and evidence
- NVIDIA Snapshot: GLM-5.3 restored in 31 s on NVIDIA HGX B300 (10 November 2026, 00:00 UTC)
Independent WittyWires tracker for public updates about Crusoe. Not affiliated with or endorsed by Crusoe; this is not an official account.