DGX Spark storage: what's persistent and what to back up
One node, 1 TB of NVMe: stopping keeps your /workspace, terminating deletes it, and GPUwerk keeps no backup for you. This page lays out exactly what survives which action, using only what GPUwerk's pricing and SSH docs already state.
The storage itself
Every GPUwerk DGX Spark node (on-demand or in the 256 GB cluster) comes with 1 TB of NVMe, per the spec tables on the pricing page and the hardware page. You land as root in /workspace, in the console's browser terminal or over SSH if you add your own key, per GPUwerk's SSH keys documentation. That 1 TB is shared between the container's own filesystem and whatever you keep in /workspace. How many current models fit in 1 TB or 4 TB, and how to share one model cache between engines, is in DGX Spark model storage.
The one directory that matters: /workspace
/workspace is the part of the container that survives a stop. Per the pricing FAQ, stopping keeps it right where it is, on the same node. Anything installed outside /workspace lives only on the container filesystem and is gone after a rebuild, per the SSH documentation. The practical rule: keep your models, checkouts and datasets under /workspace, and treat anything installed elsewhere as disposable setup you'd need to redo.
Stop versus terminate, exactly
These two actions do different things to your data, and the pricing page's FAQ spells out both precisely:
- Stop keeps
/workspacein place on the same node. The Spark stays held for you, the container stops, and the SSH port and hostname don't change, so restarting resumes the same machine, though model presets may need several minutes to warm back up. A customer-requested stop bills 75% of the running rate to keep the Spark held for you: a 25% discount off the $0.79/hour on-demand rate. - Terminate deletes your workspace from the node for good, which is why the console requires typing the instance name to confirm. This is the way to stop paying entirely; a stopped instance keeps billing at the reduced rate as long as its Spark is held.
There's a third path worth naming separately, because it isn't something you choose: if your credit reaches zero or you hit a monthly budget, the instance is stopped and its machine released. The workspace stays on that machine for up to 48 hours. Top up (or raise the budget) and press start in that time to pick up where you left off, unless someone else has rented the machine, which deletes it sooner. After 48 hours it is deleted for good, per the pricing page, and nothing is saved off the machine. Low-balance emails warn you before this happens, and auto top-up can reduce the risk, but the pricing page notes plainly that payments can fail, so a low-balance warning is not a guarantee you'll catch it in time.
No backup service, stated plainly
GPUwerk's pricing page states this without qualification: workspaces are not backed up. GPUwerk keeps no copy of /workspace at all, not even one for a failed drive, so if the drive fails, the workspace is lost. The pricing page's own callout on this point is direct: "Protect your results... Keep your own backup."
In practice, anything in /workspace you can't afford to lose needs its own copy somewhere else before you terminate, ideally on a schedule if the instance runs long enough that a zero-balance stop is a real risk. rsync, scp, or pushing to your own object storage or git remote over the SSH tunnel documented in the SSH keys guide all work for this. GPUwerk doesn't provide or manage the backup destination; that choice is yours.
Practical checklist
- Put everything worth keeping in
/workspace. Nothing outside it survives a rebuild. - Stop, don't terminate, if you want the exact same machine back later at a 25% discount off the running rate.
- Terminate only once you've pulled anything you need, since it deletes the workspace immediately and irreversibly.
- Keep your account funded, or set up auto top-up: after a zero-balance stop the workspace stays on its machine for up to 48 hours, or less if someone else rents the machine, and then it's deleted the same as a manual terminate, with no backup on GPUwerk's side either way.
- Back up anything genuinely important off the node, on your own schedule, using your own storage.
See the full billing FAQ for the exact stop and terminate pricing, or a first engagement if you want help designing a backup and rollout plan around a specific workload.