Network & storage
Storage and backups
Local NVMe on DediGPU servers: capacities per card, extra NVMe, RAID-1, what happens to data on suspension and cancellation, and why you need your own backups.
Storage is local NVMe in the chassis, sized per card. It is the fastest thing you can rent and it is also a single device, which is why the rest of this page is about copies.
Included NVMe
| Card | Per card | 8-card node |
|---|---|---|
| B300 | 3 TB | 24 TB |
| B200 | 2 TB | 16 TB |
| H200 | 2 TB | 16 TB |
| H100 SXM | 1.5 TB | 12 TB |
| H100 PCIe | 1 TB | 8 TB |
| MI300X | 2 TB | 16 TB |
| A100 80 GB | 1 TB | 8 TB |
| RTX PRO 6000 | 1 TB | 8 TB |
| L40S | 750 GB | 6 TB |
| RTX 6000 Ada | 750 GB | 6 TB |
| RTX A6000 | 500 GB | 4 TB |
| RTX 5090 | 400 GB | 3.2 TB |
| RTX 4090 | 300 GB | 2.4 TB |
Extra NVMe
$6 per TB per month, in 2, 4 or 8 TB, added as a separate device (/dev/nvme1n1). Ask for RAID-1 in the order note and the two devices are mirrored at no charge. Enterprise drives rated for sustained writes, 7 GB/s sequential per device.
What happens to your data
- Reboot, re-provisioning of a template
- Untouched.
- Re-image
- Wiped: the OS and template are reinstalled on a clean device.
- Suspension (balance short)
- Kept 14 days, powered off. Resume and it is all there.
- Cancellation
- The devices are erased with the drive's secure-erase command before the hardware is reused. Immediately.
- Drive failure
- The device is replaced from spares within the SLA target. Data on the failed device is gone.
Backups
We keep none, by design: we do not want a copy of your data, and the cost sheet does not carry one. Keep checkpoints and datasets somewhere else. Egress is free, so a nightly rclone sync or restic backup to an object store of your choice costs you only the store.
restic -r s3:s3.example.net/checkpoints init
restic -r s3:s3.example.net/checkpoints backup /data/checkpoints
# in cron, every night