Interactive tool
AWS EFS cost calculator for ML training
Drag the sliders to match your workload. See your AWS EFS monthly bill next to your Training Pipes equivalent — and how much switching would save you.
Your workload
Tune these sliders to match your training pipeline.
Total unique training data in object storage.
Percentage of the dataset read in any given week.
Full dataset reads per month across all training jobs.
Checkpoints and artifact writes per month.
Your monthly bill
Based on list pricing (us-east-1) for EFS Elastic Throughput and Training Pipes plans as of today.
- Standard$2,250
- IA$1,063
- Reads$60,000
- Writes$120.00
- Object storage$750.00
- Scale plan$999.00
How the math works
EFS Elastic Throughput charges two things: provisioned capacity (at Standard or IA rates depending on access frequency) and per-GB request costs for reads and writes. For a dataset that's read many times per month — which is most ML training — the per-request charges dominate. Training Pipes uses cheap object storage as the durable tier and a regional NVMe cache for hot reads. Repeated reads of cached data don't incur per-read charges, which is where the savings come from.
See the full cost breakdown for worked examples and the EFS vs Training Pipes comparison for a feature-by-feature view.
Frequently asked questions
- What pricing does this calculator use?
- AWS EFS us-east-1 list prices for Elastic Throughput ($0.30/GB Standard, $0.025/GB IA, $0.03/GB read, $0.06/GB write). Training Pipes uses a flat per-GB object storage price ($0.015/GB) plus a plan tier.
- What is the "hot working set"?
- The percentage of your dataset that is read in a typical week. For most ML training workloads this is 10-30% — the rest is cold between epochs.
- Does this include cross-region egress?
- No. Both calculations assume the storage is in the same region as your compute. Cross-region egress applies separately to both systems.
- Why is Training Pipes so much cheaper for ML?
- EFS charges by provisioned capacity. ML training workloads read the same data many times but most data is cold between epochs, so provisioned-capacity pricing is expensive. Training Pipes pairs cheap object storage for cold data with a regional NVMe cache for the hot set.