Calculator inputs
Use the form to estimate machine learning epoch costs with compute, energy, storage, labor, and overhead inputs.
Example data table
These examples show how different training profiles can change total spending and average epoch cost.
| Scenario | GPUs | Epochs | Effective hours per epoch | Total cost | Average cost per epoch |
|---|---|---|---|---|---|
| Vision fine-tune | 4 | 20 | 2.12 | $831.74 | $41.59 |
| Small language adaptation | 2 | 10 | 1.00 | $117.49 | $11.75 |
| Large distributed training run | 8 | 30 | 3.13 | $3,473.37 | $115.78 |
Formula used
1) Steps per epoch
Steps per epoch = ceil(training samples / batch size)
2) Effective epoch time
Effective epoch time = base epoch time / (utilization % / 100)
3) Compute cost
Compute cost per epoch = hourly GPU rate × GPU count × effective epoch time × (1 - discount %)
4) Power cost
Power kWh per epoch = power watts per GPU × GPU count × effective epoch time × PUE / 1000
Power cost per epoch = power kWh per epoch × electricity rate
5) Storage cost
Total storage cost = storage GB × storage rate per GB-month × (training days / 30)
6) Egress and labor cost
Total egress cost = data egress GB × egress rate per GB
Total engineer cost = engineer hours × engineer hourly rate
7) Final run cost
Subtotal = compute + power + storage + egress + labor
Grand total = subtotal + (subtotal × overhead %)
Average cost per epoch = grand total / total epochs
How to use this calculator
- Enter a project name so exported reports stay easy to identify.
- Provide training samples, batch size, planned epochs, and the base epoch duration.
- Fill in hardware pricing, GPU count, discount, and utilization values.
- Add energy, storage, egress, labor, and overhead assumptions.
- Press Calculate epoch cost to see results above the form.
- Review the breakdown table and Plotly graph for budget analysis.
- Use the CSV or PDF buttons to export the current report.
Frequently asked questions
1) What is epoch cost?
Epoch cost is the spending required to complete one full pass through your training dataset. It combines compute rental, electricity, storage, labor, transfer fees, and overhead into a single planning number.
2) Why does utilization affect epoch time?
Low utilization means your expensive hardware is not fully busy. The calculator increases effective epoch time when utilization drops, which raises billed compute time and usually increases power cost too.
3) Should I include engineer hours?
Yes, when people actively monitor, tune, or troubleshoot the run. Labor may dominate small experiments, while compute may dominate large distributed training jobs. Including both gives a more realistic budget.
4) What does PUE mean here?
PUE stands for power usage effectiveness. It adjusts IT power to reflect cooling and facility overhead. A higher PUE increases the effective power consumed by the training infrastructure.
5) Is storage charged per epoch?
Not directly. The calculator converts total runtime into days, then estimates storage using a monthly rate. That time-weighted cost is distributed across epochs for reporting clarity.
6) How should I estimate discount percent?
Use the expected savings from reserved, committed, or spot pricing compared with the standard on-demand rate. Enter zero if your hardware is billed without any discount program.
7) Can I compare multiple hardware setups?
Yes. Run the calculator several times with different GPU counts, rates, and utilization values. Average cost per epoch and total cost make side-by-side scenario comparison much easier.
8) What is cost per 1,000 samples useful for?
It normalizes spending across datasets and training plans. Teams often use it to compare experiments with different epoch counts, sample sizes, or model scales using one comparable metric.