Optimizing Data Transfer Costs Simplifies | Generated by AI
Yes, moving data across regions costs time (network transfer). DO doesn’t have a “hyper link” between regions — it goes over public internet, roughly 100-500 Mbps, so 100 GB = ~30 min to a few hours.
But honestly, the simplest approach: just recreate the GPU droplet when ready, download and prep data directly on it. Data download + tokenization for GPT-2 training data is maybe 1-2 hours of GPU time = ~$4-8. Not worth optimizing around.
The real cost is idle time, not compute time.
Here’s the math:
| Approach | Cost |
|---|---|
| GPU droplet idle for 1 week while prepping data on CPU | $336 GPU idle + $16 CPU |
| GPU droplet on for 2 hrs to download + tokenize, then train | $4-8 for prep + training cost |
Recommendation: Keep things simple.
- Delete the atl1 volume ($10/mo saved) — you have no data on it
- Delete all snapshots except
gpu-mi300x-snapshot-20260601($7/mo saved) - When ready to train: recreate GPU droplet from snapshot, download data from HuggingFace, tokenize, train
- Total idle cost: ~$1.53/mo (just the snapshot)
Want me to clean up and go this route?
Generated by AI. Curating and sharing still takes effort.
If you find it useful, feel free to
donate.
WeChat: @lzwjava
·
X: @lzwjava ·
Say hi 👋